Representation and Mind

Hilary Putnam and Ned Block, editors

Represmtation and Reality Hilary Putnam
Explaining Behavior: Reasons in a World of Causes Fred Dretske
The Metaphysics of Meaning Jerrold J. Katz
A Theory of Content and Other Essays Jerry A. Fodor
The Realistic Spirit: Wittgenstein, Philosophy, and the Mind Cora Diamond
The Unity of the Self Stephen L. White
The Imagery Debate Michael T ye
A Study of Concepts Christopher Peacocke
The Rediscovery of the Mind John R. Searle
Past, Space, and Self John Campbell
Mental Reality Galen Strawson
Ten Problems of Consciousness: A Representational Theory of the Phenomenal
Mind Michael Tye
Representations, Targets, and Attitudes Robert Cummins
Starmaking: Realism, Anti-Realism, and frrealism Peter J. McCormick (ed.)
A Logical Journey: From Godel to Philosophy Hao Wang
Brainchildren: Essays on Designing Minds Daniel C. Dennett
Realistic Rationalism Jerrold J. Katz
The Paradox of Stlf-ConsciC1USness Jose Luis Bermudez
In Critical Condition: Polemical Essays on Cognitive Science and the Philosophy
of Mind Jerry Fodor
In Critical Condition
Polemical Essays on Cognitive Science and the
Philosophy of Mind
Jerry Fodor
A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England
First MIT Press paperback edition. 2000
1998 Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any electronic or
mechanical means (including photocopying. recording. or information storage and retrieval)
without permission in writing from the publisher.
This book was set in Palatino on the Monotype "Prism Plus" PostScript lmagesetter by Asco
Trade Typesetting Ltd .. Hong Kong.
Printed and bound in the United States of America.
Library of Congress Cataloging-in-Publication Data
Fodor. jerry A.
In critical condition : polemical essays on cognitive science and
the philosophy of mind I Jerry Fodor.
p. em. - (Representation and mind)
"A Bradford book."
Includes bibliographical references (p. ) and index.
ISBN: 9780262-$6128-0 : alk. paper). 0-262-56128-X (pb)
1. Philosophy of mind. 2. Cognitive science. I. Iitle.
II. Series.
BD418.3.F625 1998
128'.2-dcll
98-21611
CIP
Contents
Preface ix
Part I
Metaphysics 1
Chapter 1
Review of John McDowell's Mind and World 3
Chapter 2
Special Sciences: Still Autonomous after All These Years (A Reply to
Jaegwon Kim's "Multiple Realization and the Metaphysics of
Reduction") 9
Part II
Concepts 25
Chapter 3
Review of Christopher Peacocke's A Study of Concepts 27
Chapter 4
There Are No Recognitional Concepts-Not Even RED 35
Chapter 5
There Are No Recognitional Concepts-Not Even RED, Part 2: The Plot
Thickens 49
Chapter 6
Do We Think in Mentalese7 Remarks on Some Arguments of Peter
Carruthers 63
Chapter 7
Review of A. W. Moore's Points of View 75
Part III
Cognitive Architecture 81
viii Contents
Chapter 8
Review of Paul Church land' s The Engine of Reason, The Seat of the
Soul 83
Chapter 9
Connectionism and the Problem of Systematicity: Why Smolensky's
Solution Doesn't Work 91
jerry Fodor and Brian McLaughlin
Chapter 10
Connectionism and the Problem of Systematicity (Continued): Why
Smolensky's Solution Still Doesn't Work 113
Chapter 11
There and Back Again: A Review of Annette Karmiloff-Smith's Beyond
Modularity 127
Chapter 12
Review of Jeff Elman et al.. Rethinking Innateness 143
Chapter 13
Review of Steven Mithen's The Prehistory of the Mind 153
Part IV
Philosophical Darwinism 161
Chapter 14
Review of Richard Dawkins's Climbing Mount Improbable 163
Chapter 15
Deconstructing Dennett's Darwin 171
Chapter 16
Is Science Biologically Possible? Comments on Some Arguments of
Patricia Churchland and of Alvin Plantinga 189
Chapter 17
Review of Steven Pinker' s How the Mind Works and Henry Plotkin's
Evolution in Mind 203
References 215
Index 217
Preface
Philosophy, like piloting, is mostly about 6guring out where you are. The
basic principle of both is much the same: Find an object whose position is
known and locate yourself with respect to it. Typically, in both cases, the
object in question is somewhere that you do not wish to be: On a rock,
on a shoal, at the edge of the channel, or half a mile inland from the shore.
So the trick is to get close enough to recognize the thing and to 6gure out
what it means, but not so close that it swallows you up. Thus Aristotle
with respect to Plato, Kant with respect to Hume, Descartes with respect
to Newton, and me with respect to many a bell buoy on many a summer
Sunday afternoon.
The difference is that whereas navigators don't often argue with their
landmarks, philosophers hardly ever do anything else. Hence this collec-
tion. Each of these essays reacts to a position in .philosophy or cognitive
science (in the present context the distinction is hardly worth drawing)
that I'm pretty sure I don't want to occupy, but that I'm also pretty sure
needs to be marked on the charts and, as sailors say, ''honored." I think
that not getting wrecked requires avoiding all of them at once. And I
think that there's a way to do that: What's required is a mix of intentional
realism, computational reductionism, nativism, and semantic atomism,
together with a representational theory of mind. I've talked and written
about all that elsewhere, however; in these polemical essays the positive
program is present largely by implication. Very roughly, and just for ori-
entation: The views examined in Part I bear on questions of realism and
reductionism, and those in Part II mostly concern semantic atomism. Part
III is about what, if any, sense can be made of the idea that mental pro-
cesses are computational. Part IV is about what, if any, constraints evolu-
tionary theory imposes on semantics and the philosophy of mind.
Many of these papers are book reviews and the targets are identi6ed
by the titles. Others are responses to journal articles which are likewise
eponymous. The apparent exceptions are chapters 4 and S, the ones
about recognitional concepts. In fact, however, these were prompted by,
and constitute an extended cornrnentary on, a philosophical project that
x Preface
Christopher Peacocke has been working out for some years and which is
most explicitly set forth in his book A Study of Concepts (see chapter 3 of
the present volume for an overview). Quite a lot of the philosophy I've
done of late is an attempt to not be co-opted by Peacocke's project, while
at the same time not getting lost. It turns out that's not so easy. I'm
indebted to Peacocke for having provided a particularly bright light to
steer by-and away from.
Chapter 1 was originally published in The London Review of Books, April
20, 1995. Chapter 2 was originally published in}. Tomberlin (ed)., Philo-
sophical Perspectives, 11, Mind, Causation, and World, Ridgeview, CA.
Chapter 3 was originally published in The London Review of Books, April
20, 1995. Chapter 4 was originally published in Philosophical .Issues, Vol. 9
(E. Villanueva, ed.), Ridgeview Publishing Co., 1998. Chapter 7 was origi-
nally published in The London Review of Books, October 30, 1997. Chapter
8 was originally published in The Times Literary Supplement, August 25,
1995. Chapter 9 was originally published in Cognition, 1990,35, 183-204.
Chapter 10 was originally published in Cognition, 1996, 62, 109-119. (I'm
grateful to Professor Brian McLaughlin for agreeing to my reprinting this
paper.) Chapter 12 was originally published in The London Review of
Books, Vol. 20, No. 12, 1997. Chapter 13 was originally published in The
London Review of Books, Vol. 18, No. 23, 1996. Chapter 14 was originally
published in The London Review of Books, Vol. 18, 1996. Chapter 15 was
originally published in Mind and Language, 1996, Vol. 11, No.3, pp. 246-
262. Chapter 17 was originally published in The London Review of Books,
Jan. 22, 1998. I am grateful to all of these sources for their kind permis-
sion to publish this material here.
A word about style: many of these papers were originally intended for
an audience that was either lay or interdisciplinary. In those, hard core
philosophy is kept to a minimum. There are also a lot of jokes, for which I
am sometimes taken to task. I do admit to finding many of the views that
I argue against pretty funny. I find many of my own views pretty funny
too. Nietzsche is righter than Brecht: Sometimes the man who laughs has
heard the terrible news.
Part I
Metaphysics
Chapter 1
Review of John McDowell's Mind and World
Mind and World collects McDowell's 1991 John Locke lectures, with
some of his afterthoughts appended. Until you're halfway through, the
book seems to be about a relatively technical question: What's the rela-
tion between the mind's contribution to perceptual experience and the
world's? A deep worry, to be sure; but one that's maybe best left for epis-
temologists to thrash out. In chapter 4, however, McDowell achieves a
coup de theatre. The angle of view widens unexpectedly, and the specifi-
cally epistemological issue is seen to be a microcosm of some of modem
philosophy's most characteristic concerns: How should we construe the
relation between the realm of reason and the realm of natural law? How
can it be that we. are both physical objects and thinking things? How can
conceptualization be free and spontaneous if the mind is a mechanism?
Why, in short, isn't "rational animal" an oxymoron? This is a fine exposi-
tory contrivance, and the framing of the issues that it permits is deeply
illuminating. McDowell's book finds new ways to register a number of
our deepest perplexities.
Which, however, is not to say that I believe a word of it.
A dialectic between two different and opposed conceptions of natural-
ism-in particular, of a naturalistic account of rationality-is working
itself out in Mind and World. There's the reductionist version (McDowell
calls. it ''bald" naturalism; "scientism" is another pejorative that is cur-
rently in fashion). And there's the kind of naturalistic pluralism that
McDowell himself is striving for. Very roughly, the distinction is between
the tradition that runs &om Kant through the positivists to the likes
of Dewey and Quine, and the tradition that runs &om Kant through the
Hegelians to Wittgenstein, Ryle, and Davidson (and Hilary Putnam since
he left MIT for Harvard).
It's hard to be articulate about this disagreement; we're very close to the
edge of what we know how to talk about at all sensibly. But, for reduc-
tionists, the world picture that the natural sciences lay out has a sort
of priority-sometimes viewed as metaphysical, sometimes as methodo-
logical, sometimes as ideological, sometimes as all of these at once-to
4 Chapter 1
which other discourse is required to defer insofar as it purports to speak
literal truths. Conflicts between the scientific image and, for example, the
claims that moral theories make, or theories of agency, or theories of
mind, are real possibilities, and if they arise, science is privileged-not
because the "scientific method" is infallible, but because the natural realm
is the only realm there is or can be; everything that ever happens, includ-
ing our being rational, is the conformity of nature to law. And our science
is the best attested story about the conformity of nature to law that we
know how to tell. Accordingly, the philosophical problems about mind
and world have to be situated within the general scientific enterprise, if
literal truth is what philosophers aim for. What our rationality consists in
is an open question, apt for a kind of inquiry that is empirical and meta-
physical at the same time; as, indeed, scientific inquiry is wont to be.
For pluralists, however, the situation presents itself quite differently.
There are lots of more or less autonomous varieties of discourse (or world
views, or language games, or forms of life, or paradigms), and the critique
they are subject to is largely from inside and in their own terms. As
McDowell puts it, "Even a thought that transforms a tradition must be
rooted in the tradition that it transforms" {187; all references in this paper
are to McDowell, 1994). Accordingly, the natural scientist's activity of
limning a normless and otherwise "disenchanted" natural order is just one
way of world-making among others. For the epistemologist's purposes, in
contrast to the scientist's, the normative character of rational assessment
is given a priori; to that extent, we already know what the essence of
rationality is. The problem is to find a place for it outside what the natural
sciences take to be the natural order, but to do so without, as McDowell
sometimes says, thereby making rationality look spooky.
Pretty clearly McDowell thinks that bald, reductive naturalism isn't
seriously an option, so it's going to be pluralism or nothing if the integ-
rity of the rational is to be sustained. Just why he thinks this is less clear.
Circa page 7 4 there's some moderately loose talk, with a nod to Donald
Davidson, about the "constitutive principles" of rationality being such
that "the logical space that is the home of the idea of spontaneity cannot
be aligned with the logical space that is the home of ideas of what is
natural in the relevant sense [viz. with] the characteristically modem con-
ception according to which something's way of being natural is its posi-
tion in the realm of law." The unwary reader might suppose from this that
somebody has actually argued conclusively that the reductionist program
can't be carried through, hence that either the mutual autonomy of the
natural order and the rational order is somehow guaranteed, or else there
is no such thing as rationality; the latter outcome being, to use a favorite
epithet of McDowell's, "intolerable."
Review of McDowell's Mind and World 5
In fact, of course, nothing of the sort has been shown, nor will be. Bald
naturalism may, for all philosophers know, be viable after all. That may
strike you as comforting if, like me, you find McDowell's efforts to
formulate a pluralistic alternative rather less than convincing.
I guess I am a hairy naturalist: Though I do agree that the problems
about mind and world are a lot harder than reductionists have sometimes
supposed, I also think that an adequate and complete empirical psychol-
ogy would, ipso facto, tell the whole, literal truth about the essence of
the mental. Science discovers essences, as Kripke once remarked. So, if
it's literally true that rationality, intentionality, normativity, and the like
belong to the mind essentially, then they must all be phenomena within
the natural realm that scientists explore. McDowell comments, sort of in
passing, that "cognitive psychology is an intellectually respectable disci-
pline ... so long as it stays within its proper bounds." That, however, is
truistic, proper bounds being what they are. The serious question is
whether there is anything about mentality that can be excluded from
scientific inquiry a priori; anything, anyhow, that claims to be both of the
essence and literally true.
Consider, for example, the epistemological question that McDowell
starts with. You might have thought that seeing a tree goes something
like this: Light bounces off the tree and affects your eyes in ways that
determine the sensory content of your experience. Your mind reacts by
inferring from the sensory content of your experience something about
what in the world must have caused it. The upshot, if all goes well, is that
you see the world as locally entreed.
So, then, perception is a hybrid of what the senses are given and what
the mind infers. The process is causal through and through: It's part of
psychophysics that encounters with trees bring about the kinds of visual
sensations that they do. And it's part of cognitive psychology that the
visual sensations that encounters with trees provoke occasion the percep-
tual inferences that they do in minds with the right kinds of history and
structure. To be sure, the story as I've just told it is bald and insufficiently
detailed; but ironing out its wrinkles is what perceptual psychologists are
paid to do, and my impression is that they're getting along with the job
pretty well. So what, exactly, is supposed to be wrong?
McDowell's answer is entirely characteristic: ''The trouble about the
Myth of the Given is that it offers us at best exculpations where we wanted
justifications .... [T]he best [it] can yield is that we cannot be blamed for
believing whatever [our experiences] lead us to believe, not that we are
justified in believing it" (13). That is: McDowell has in mind a certain
account of what the rationality of a perceptual judgment consists in, and
that account isn't satisfied if the world's contribution is merely to provide
the judgment with the right sort of etiology. The empirical story says that
6 Chapter 1
what the world gives beliefs come from the sensory states that it causes
us to have. But that won't do since sensations aren't reasons; "nothing can
count as a reason for holding a belief except something else that is also
in the space of concepts ... " (140). Period. To have to tinker with the
epistemological story-to have to reconsider what rational judgment is
in light of a merely empirical theory about the conditions under which it
can be achieved-would strike McDowell as, well, intolerable.
So McDowell is committed to a view that might well strike you as
hopeless on the face of it; he needs to be a naturalist and a dualist at the
same time. On one side, what the world contributes to perception must be
something that one can think. It must have, so to speak, the kind of struc-
ture that thoughts have, for only then does the mental process that gets
from experience to judgment count as rational by McDowell's criteria.
But, on the other side, McDowell is quite aware that the world isn't any
kind of text, and that idealism has to be avoided. '1n a common medieval
outlook, what we now see as the subject matter of natural science was
conceived ... as if all of nature were a book of lessons for us .... It is a
mark of intellectual progress that educated people cannot now take that
idea seriously'' (71). One is therefore not to argue as idealists might wish
to do: that if the world's contribution to perception has to be something
that's thinkable, then since all that is thinkable is thoughts, the world must
be made of thoughts if we are to be able to perceive it.
These views are, however, alternatives that a lot of philosophers hold
to be exhaustive. Granting something unconceptualized that is simply
Given to the mind in experience has generally been supposed to be the
epistemological price one has to pay for an ontology that takes the
world to be not itself mind-dependent. The epis-temological passages in
McDowell's book struggle to find space between these options. I'm not
at all sure they do. I'm not at all sure that there is any space to find. From
the comfortable perspective of my kind of naturalist, anyhow, the whole
undertaking seems to carry contortion beyond necessity. Laocoon looks a
little comical if there's no snake. Until the second appendix, McDowell
doesn't even consider taking what is surely the easy way out: Maybe
sometimes exculpation is justification and is all the justification that there
is to be had.
Why not, after all? If the situation is that I can't but believe that I'm
looking at a tree, and if, in that situation, it's the case that I am looking at
a tree, and if there is a workable account of why, in such situations, I reli-
ably come to believe that there's a tree that I'm looking at (viz., because
they are situations where trees cause the kind of sensations that cause
minds like mine to think that they are seeing trees), why isn't that good
enough for my judgment that I'm seeing a tree to make it into the Realm
of Reason? Why, in short, mightn't Aeshing out the standard psycho-
Review of McDowell's Mind and World 7
logical account of perception itself count as learning what perceptual jus-
tification amounts to7
When McDowell does finally consider this kind of option, he switches
ground disconcertingly. Suddenly, the worry isn't that causation provides
for exculpation but not for justification; rather, it's that the justification
it provides for needn't constitute "a subject's reasons for believing some-
thing" (sic, 161). This all goes very fast, and I doubt that it actually
amounts to much. When, in the situation imagined, I come to believe that
I see a tree, my reason for believing there's a tree is that I see it; and my
reason for believing that I see it is that I do. Why do I think I see a tree
there7 Well, why on earth should I not7 Just look at the thumping great
thing! Mcdowell's sense of how justification actually goes occasionally
strikes me as excessively earnest.
This easy path having once been eschewed, the hard path proves to
be very hard indeed: '1f we can rethink our conception of nature so as
to make room for spontaneity, even though we deny that spontaneity is
captured by the resources of bald naturalism, we shall by the same token
be rethinking our conception of what it takes for a position to be called
'naturalism'" (77). More bluntly: The cost of McDowell's a priorism is
that he has to be some sort of dualist; not necessarily the Cartesian sort,
who thinks that there are immaterial things. But, quite likely, the kind of
faculty dualist who is, willy-nilly, landed with occult powers. Having sit-
uated the rational (and the ethical, and a lot else.that we care about) out-
side the realm of law, McDowell needs to face the embarrassing question
how, by any natural process, do we ever manage to get at it7
He needs to, but in fact he doesn't. ''When we are not misled by expe-
rience, we are directly confronted by a worldly state of affairs itself, not
waited on by an intermediary that happens to tell the truth" (143). But if
you want to hold that states of affairs themselves are what veridical per-
ception works on, you need a story about how unmediated cognitive
connections to states-of-affairs themselves might be achieved. Likewise,
and more so, in the nonperceptual cases, where the objects of cognition
are normative or otherwise intentional aspects of things: how do we get
at those if they aren't in the natural order7 Maybe better, how do they get
at us7 How can what is not in the realm of law make anything happen?
Here's what I take to be McDowell's answer: 'We need to bring respon-
siveness to meaning back into the operations of our natural sentient
capacities as such" (77); as he sometimes puts it, we need somehow to
think of the mind as "resonating" to rational relations. Or consider: "the
rational demands of ethics are not alien to the contingencies of our life
as human beings .... [O]rdinary upbringing can shape the actions and
thoughts of human beings in a way that brings these demands into view"
(83). But "bringing into view" is a metaphor; only what is in nature can
8 Chapter 1
literally be viewed. And "resonating" is also a metaphor; only what is in
nature can be literally attuned to. The trouble with "putting responsive-
ness to meaning back into the operations of our natural sentient capacities
as such" is that nobody has the foggiest idea of how to do so unle,ss both
are contained in the natural order.
The forms of human sentience resonate, as far as anybody knows, only
to aspects of the "disenchanted" world. Mere exhortation won't fix that.
'We tend to be forgetful of ... second nature. I am suggesting that if we
can recapture that idea, we keep nature as it were partially enchanted, but
without lapsing into prescientiAc superstition or a rampant platonism"
(85). "[O]nce we allow that natural powers can include powers of second
nature, the threat of incoherence disappears" (88). Second nature is what
we get when "our Bildung actualizes some of the potentialities we are
born with; we do not have to suppose it introduces a nonanimal ingre-
dient into our constitution." But the question arises how second nature, so
conceived, could itself be natural. It's not good enough for McDowell just
to say that it is and you can get some at the Bildung store; he has to say
how it could be, short of spooks. Otherwise, why is McDowell's kind of
dualism to be preferred to Descartes's?
What McDowell has to pay for demanding that his account of percep-
tion conform to the a priori constraints that his normative epistemology
imposes is that he leaves us with no idea at all how perceiving could be
a process in the world. Since I, for one, can't imagine how a faculty could
resonate to meanings "as such," this seems to me not to be worth the
cost. I'm afraid the bottom line is there is no room where McDowell
wants to wiggle; a dualistic naturalism isn't in the cards. If that's right,
then epistemology needs to bend and McDowell will have to cool it a
little about justiAcation. JustiAcation can't require what can't happen, on
pain of there not being any; and whatever happens, happens in the realm
of law.
Ever since Descartes, a lot of the very best philosophers have thought
of science as an invading army from whose depredations safe havens have
somehow to be constructed. Philosophy patrols the borders, keeping the
sciences "intellectually respectable" by keeping them "within . . . proper
bounds." But you have to look outside these bounds if what you care
about is the life of the spirit or the life of the mind. McDowell's is as good
a contemporary representative of this kind of philosophical sensibility as
you could hope to And. But it's all wrong headed. Science isn't an enemy,
it's just us. And our problem isn't to make a place in the world for the
mind. The mind is already in the world; our problem is to understand it.
Chapter 2
Special Sciences: Still Autonomous after All These
Years (A Reply to }aegwon Kim's "Multiple
Realization and the Metaphysics of Reduction")
The conventional wisdom in philosophy of mind is that "the conventional
wisdom in philosophy of mind [is] that psychological states are 'multiply
realized' [and that this] fact refutes psychophysical reductionism once and
for all" (Kim, 1993, 309. All Kim references are to this paper except as
noted). Despite the consensus, however, I am strongly inclined to think
that psychological states are multiply realized and that this fact refutes
psychophysical reductionism once and for all. As e. e. cummings says
somewhere: "nobody loses all of the time."
Simply to have a convenient way of talking, I will say that a law or
theory that figures in bona fide empirical explanations but that is not
reducible to a law or theory of physics is ipso facto autonomous, and that
the states whose behavior such laws or theories specify are functional
states. (In fact, I don't know whether autonomous states are ipso facto
functional. For present purposes all that matters is whether functional
states are ipso facto autonomous.) So, then, the conventional wisdom in
the philosophy of mind is that psychological states are functional and the
laws and theories that figure in psychological explanations are autono-
mous.1 (Likewise, and for much the same reasons, for the laws, theories,
etc. in other "special'; [viz. nonbasic] sciences.) The present discussion
undertakes to defend this consensus view against a philosophical qualm
that is the main moral of Jaegwon Kim's paper "Multiple Realization and
the Metaphysics of Reduction."
Kim says that he's prepared to agree (at least for the sake of the argu-
ment) that:
1. Psychological states are typically multiply realized (MR); and
2. MR states are ipso facto unsuitable for reduction.
But Kim thinks philosophers haven't gotten it right about why MR
states are ipso facto unsuitable for reduction. Once they do, Kim says,
they'll see that the moral of 1 and 2 isn't, after all, that psychology is
autonomous. Rather, it's that quotidian psychological states aren't reduci-
ble because they aren't projectible. Unprojectible states are, by definition,
10 Chapter 2
not the subjects of a possible science; they aren't bona Sde kinds and they
can't appear in bona Sde nomological explanations. A fortiori, terms that
express psychological states are not available for incorporation in ''bridge
laws" or in (metaphysically necessary) property identities. This is all, of
course, contrary to what a lot of philosophers, to say nothing of a lot of
psychologists, have hitherto supposed.
Caveat: It turns out, according to Kim that some very Bnely indi-
viduated psychological states are '1ocally" reducible after all; but that's
because these states are, by assumption, not MR. More on this later; I
mention it here just to make clear that Kim is not claiming that psycho-
logical states are unprojectible qua psychological (or qua intentional), but
only that they are unprojectible qua not local. Contrast (e.g.) Davidson
(1980).
Now, I think Kim is quite right that what moral you should draw from
MR states not being reducible depends on whether whatever it is that
makes them not reducible also makes them not projectible. But I think that
the diagnosis of the irreducibility of MR states that Kim offers is wrong,
and that the right diagnosis supports the standard view: "pain," ''believes
that P," and the like express real states, about which all the available evi-
dence suggests that there are reallaws.
2
If I'm right that such states are
projectible but not reducible, then it follows that psychological laws are
autonomous after all.
So much for strategy; now for tactics. Kim's polemical method is to pick
an apparently untendentious (and, in particular, a nonpsychological) MR
state and explain why it is unprojectible qua MR. He will then argue that
since beliefs, pains, and the like are MR by assumption (1), they must be
unprojectible for the same reasons that his MR paradigm is. Tit for tat: I'll
argue both that Kim's diagnosis of the unprojectibility of his MR para-
digm is wrong and that the supposed analogy between his MR paradigm
and pains, beliefs, and the like is spurious.
More caveat: In order not to be always writing "pains, beliefs, and
the like" or "quotidian psychological states," I'll follow Kim's practice and
take pain as my working example of a quotidian psychological state. In
particular, I'll assume that, if there are any psychological laws at all, then
probably there are psychological laws about pain. In fact, however, pain
isn't really a happy choice for a working example, since though I take it to
be quite a plausible candidate for projectibility, pain is notoriously a very
bad candidate for being MR, hence for functional analysis. That, however,
is stuff for a different paper. (See, e.g., Block and Fodor, 1972; Block,
1978.) My present brief is just that what Kim says is wrong with func-
tionalism actually isn't; so the autnonomy thesis is, to that extent at least,
not in jeopardy. For these purposes, pain will do.
So, Bnally, to work.
Special Sciences 11
jade: We are told, Kim tells us, that jade "is not a mineral kind, con-
trary to what was once believed; rather, jade is comprised of two distinct
minerals with dissimilar molecular structure, jadeite and nephrite" (319).
Geological unsophisticate that I am, I shall often call these two minerals
jadeA and jadeB in the discussion that follows. It won't really matter
which is which, so you needn't bother to keep track.
Kim thinks that, because of these fads about jadeite and nephrite, jade
is paradigmatically MR. Kim also thinks that since jade is paradigmatically
MR, "jade" is ipso facto unprojedible. And, a fortiori, (3) isn't a law.
3. Jade is green
I don't actually care much whether (3) is a law or even whether "jade" is
projedible. But I am going to deny that jade is paradigmatically MR, and
I am also going to deny that "jade" is unprojectible for the reason that
Kim says it is.
So then, what, according to Kim, is wrong with (3)7 Well, "D]awlike
generalizations ... are thought to have the ... property [that] observation
of positive instances, Fs that are Gs, can strengthen our credence in the
next F's being G." In short, real laws are confirmed by their instances, but
(3), according to Kim, is not: "[W]e can imagine this: on reexamining the
records of past observations, we find, to our dismay, that all the positive
instances of (3) ... turn out to have been samples of jadeite, and none of
nephrite! If this should happen ... we would not ... continue to think of
(3) as well con6rmed .... But all the millions of green jadeite samples are
positive instances of (3) .... The reason [that (3) is not confirmed by them]
is that jade is a true disjunctive kind, a disjunction of two heterogeneous
nomic kinds which, however, is not itself a nomic kind" (320)
Now notice, to begin with, that the thought experiment Kim proposes
doesn't really make the case that he wants it to. To be sure, if we dis-
covered that all our samples of green jade are samples of green jadeA,
that would lead us to doubt that (3) is well confirmed. Not, however,
because it would be a reason to think that (3). doesn't "pass the projec-
tibility test" (320); but rather, because it would show that we've made a
sampling error in collecting the data. The point is: Anybody can make a
sampling error, whether or not the hypothesis he's trying to confirm is
projectible.
Suppose we've been considering whether oak trees shed their leaves
in winter; and suppose it turns out, on reexamining the records, that our
positive instances are all observations of oak trees on the north side of
hills. Then we would no longer think of the generalization about oak
trees losing their leaves in the winter as unambiguously well confirmed;
oak data confirm oak generalizations only if they are an unbiased sample
12 Chapter 2
of the oak population, which, on the current assumption, our data aren't.
Maybe, in the present case, the generalization that the instances really
confirm is that oak trees on the north side of hills lose their leaves in winter.
But notice that discovering a sampling error of this sort would be no rea-
son at all for doubting that oak tree is a kind. Rather, the worry would be
that maybe oak tree on the north side of a hill is a kind too. If it is, then our
data are equivocal between two perfectly okay, projectible hypotheses:
the one that goes blah, blah ... oak trees ... , and the one that goes blah,
blah . ... oak trees on the north sides of hills . ... When we discover the sam-
pling error, we regard neither generalization as unequivocally confirmed
by data that are i n s t a n c ~ of both, and this is precisely because the data are
instances of both. The sampling error means that the data are equivocal, not
that the hypotheses are unprojectible. There is, to be sure, something wrong
with (3); something that makes it not a law. But what's wrong with (3)
isn't that biased samples fail to confirm it. Biased samples don't confirm
anything.
So Kim's thought experiment shows nothing that is to the point. How-
ever, he doesn't really need the thought experiment to run his argument.
Instead, he could run his argument like this:
Although green JadeA samples tell us that jadeA is green, and green
jadeB samples tell us that jadeB is green, green jadeA samples tell us
nothing at all about the color of jadeB and green jadeB samples tell us
nothing at all about the color of jadeA. That's because, though it's true
that jadeA is green iff jadeB is green, that it is true is merely accidental
given the presumed facts about the structural heterogeneity of nephrite
and jadeite. Analogously: Since jadeite is green and grass is green, jadeite
is green iff grass is. But it doesn't follow that evidence about the color of
grass bears at all on hypotheses about the color of jadeite or vice versa.
But, now, according to the functionalist orthodoxy, pain is just like jade, isn't
it? So, then:
Why isn't pain's relationship to its realization bases, Nh, Nr, Nm
analogous to jade's relation to jadeite and nephrite? If jade turns out
to be nonnomic [i.e., not projectible] on account of its dual"realiza-
tions" in distinct microstructures, why doesn't the same fate befall
pain? After all, the group of actual and nomologically possible real-
izations of pain, as they are described by the MR enthusiasts with
such imagination, is far more motley than the two chemical kinds
comprising jade. . . . We put the following question to Fodor and
like-minded philosophers: If pain is nomically equivalent to N, the
property claimed to be wildly disjunctive and obviously nonnomic,
why isn't pain itself equally heterogeneous and nonnomic as a kind? (323;
emphasis Kim's)
Special Sciences 13
I expect this question is meant to be rhetorical, but I think that I'll
answer it all the same.
Rebuttal: There is, I think, a sort of polemical standoff at this point in
the discussion. Kim is quite right that jade is "a true disjunctive kind, a
disjunction of two heterogeneous nomic kinds which, however, is not
itself a nomic kind" (12). But, from the functionalist's perspective, to say
that jade is a disjunctive kind because jade generalizations aren't con-
firmed by their instances is to put the epistemological cart before the
ontological horse. The right story, according to the functionalist, goes the
other way around. What makes jade a disjunctive kind (in the, as we're
about to see, special sense that is germane to whether it's projectible) is
that there are no general empirical truths about jade as such; a fortiori, there
are no such truths for samples of jade, as such, to confirm. Whatever is
reliably true of jade is so either because it is reliably true of jadeA as such
or because it is reliably true of jadeB as such.
No doubt there are those of you who are suspicious of "as such's"
as such; but, in fact, none is required to make the present point. What's
needed is just the distinction between a multiply based property that is
disjunctive and a multiply based property that is disjunctively realized. To
wit: A multiply based property is disjunctive iff it has no realizer in any
metaphysically possible world that it lacks in the actual world. Jade is dis-
junctive in that the only metaphysically possible worlds for jade are the
ones which contain either jadeA or jadeB or both. By contrast, multiply
based properties that are disjunctively realized have different bases in dif-
ferent worlds. Pain is disjunctively realized in that there's a metaphysi-
cally possible, nonactual, world in which there are silicon-based pains.
That is to say, in effect, that being jade just is being jadeA or jadeB.
Whereas it would be simply question-begging of Kim to hold that being
pain is the property of being one or another of pain's realizers. Function-
alists claim that pains and the like are higher-order, relational properties
that things have in virtue of the pattern of causal interactions that they
(can or do) enter into. If so, then pains, though multiply based, are not disjunc-
tive but MR. To repeat: Though Kim says that he concedes that psycho-
logical properties are MR, that's only because he isn't distinguishing being
MR (like pain) from being disjunctive (like jade). But it's exactly the dis-
tinction between disjunctiveness and disjunctive realization that function-
alists are insisting on when they say that pain states are nomologically
homogeneous under their functional description despite the physical hetero-
geniety of their realizers. You can't (and Kim can't) refute this claim just by
defining "disjunctive kind" so that it isn't true.
This is, as I say, a sort of polemical standoff. The functionalist assumes
that there are laws about pains "as such"; so he infers that, though pain is
14 Chapter 2
multiply based, it is not (merely) disjunctive. So he infers that pain is
unlike jade in the respects that are relevant to the question of projec-
tibility. Kim, going the other way around, assumes that pain is (merely)
disjunctive, hence that it is relevantly similar to jade, hence that t h ~ r e
aren't any laws about pain. The real issue-the one that Kim begs with
his appeal to the jade example-is whether there is a difference between a
multiply based property's being disjunctive and its being MR; and, if so,
what the difference is and whether it matters to projectibility. Kim almost
sees this in the closing sections his paper. But then he gets it wrong-
fatally, in my view. '1s jade a kind? We know it is not a mineral kind; but
is it any kind of a kind .... There are certain shared criteria, largely based
on observable macroproperties of mineral samples (e.g. hardness, color,
etc.), that determine whether something is a sample of jade .... What all
samples of jade have in common is just these observable macrophysical
properties that define the applicability of the predicate 'is jade'" (24). This,
I say, is just wrong; and resolving the metaphysical issues about projec-
tibility that Kim has raised turns on seeing that it is.
Suppose that, puttering around in the cellar one day, you succeed in
cooking up a substance-out of melted bottle glass, let's say-that is, for
all macroscopic purposes, indistinguishable &om jade: For example, it's as
similar in color to jadeA and jadeB as they typically are to one another; its
hardness falls at about the right place between talc and diamond on the
scratch test; it cracks along the right sort of cleavage planes; it weighs
about the same as jade per unit volume, and so forth. Have you, then,
created jade? Oh &abjous day! Oh joy that alchemists never knew! Oh (in
particular) riches running wild!
Not on your Nelly. What you've got there isn't jade; it's just melted
bottle glass. Melted bottle glass maybe counts as artificial jade in the sort
of case that we've imagined; but do not try to sell it as the real stuff. You
will find, if you do, that fanciers of jade are not amused. They will call you
unkind things like thief and fraud; and if they catch you, they will lock
you up. Pace Kim, being jade is not relevantly like having a functional
(i.e., MR) property; if it were, you could make new jade by mimicking
the macroscopic properties that jadeA and jadeB share (the ones which,
according to Kim, "determine whether something is a sample of jade, or
whether the predicate 'is jade' is correctly applicable to it"). But you can't
make jade that way. If you want to make some jade, you have to make
either some jadeite or some nephrite; just as, if you want to make some
water, you have to make some H20. That's because jade is jadeite or neph-
rite is metaphysically necessary, just like water is H
2
0.
3
As with most of
the metaphysical claims one comes across these days, the one that I just
made relies for its warrant on a blatant appeal to modal intuitions. But I
think the modal intuitions that I'm mongering are pretty clearly the right
Special Sciences 15
ones to have. If you don't share them, perhaps you need to have yours
looked at.
Now compare pain according to the functionalist persuasion. Suppose
you should happen, one day down in the cellar, to throw together a
robot, among whose types of internal states there's one that is, under
functional description, about as similar to my pains as my pains are to
yours. Then, functionalists say, the thing that you've created has pains.
Not artificial pains, not pain simulations, not virtual pains, but the real
things. Even if, as we may suppose, what you've cooked up is a silicon
robot that's made out of melted bottle glass. My point is not, I emphasize,
to claim that the functionalist is right to be of this persuasion. For reasons
I've elaborated elsewhere, I doubt that pain is a functional kind, or that
beliefs-that-P are either. What's at issue, remember, isn't whether func-
tionalism is true of the mental; it's only whether there is a plausible func-
tionalist response to the challenge that Kim has laid down: ''Why isn't
pain's relationship to its realization bases ... analogous to jade's relation-
ship to jadeite and nephrite?" Reply: There's a difference between being a
functional property (being multiply realized) and being a disjunctive prop-
erty. Being jade, according to the geologists, is an example of the latter;
being pain, according to the psychologists, is an example of the former.
So there is, thus far, nothing to warrant an inference from the unpro-
jectibility of "jade" to the unprojectibility of "pain." Kim is quite right to
emphasize that "the group of actual and nomologically possible [and epis-
temically and metaphysically possible] realizations of pain, as they are de-
scribed by the MR enthusiasts with such imagination is far more motley
than the two chemical kinds compromising jade." What he's missed is
that for that very reason the pain/jade analogy is flawed and won't do the
polemical work that he wants it to.
If there is after all plausibly a difference between the relation between
pain and its realizers, on the one hand, and the relation between jade and
its realizers, on the other, then it's not patently irrational of a functionalist
to agree that "jade" isn't projectible while continuing to claim that "pain"
is. But what is not patently irrational may nonetheless be unreasonable
and obstinate. Granted that pain and jade differ in the ways that we have
-been exploring, it remains open to Kim to wonder why that kind of
difference should matter so much. How is it that MR properties are pro-
jectible but disjunctive properties aren't? Or, to put the same question
slightly differently: Functionalists are required to deny that pain is identi-
cal to the disjunction of its realizers. The reason functionalists are required
to deny this is that it's part of their story that the property realized, but
not the disjunction of its realizers, is projectible. And the reason function-
alists have to say that is that otherwise multiple realization wouldn't be
an argument against reduction: What is supposed to make the case for the
16 Chapter 2
autonomy (unreducibility) of functional laws is that there are no laws
about the disjunction that realizes a functional state even if there are laws
about the functional state itself. So then functionalists must themselves think
that the disjunction of realizers of even bona fide, projectible MR states
aren't themselves projectible. What justifies a functionalist in making this
claim?
Or, to put the question yet another way: Functionalists hold that the
biconditionals that connect functional properties with their realizers aren't
laws (a fortiori, that they aren't ''bridge" laws). They can't be laws because
the realizers of functional states are, by assumption, disjunctive; and dis-
junctive properties are supposed to be ipso facto not projectible. Thus, in
one of his earliest functionalist papers, Putnam remarks that rendering
realizers projectible by "defining the disjunction of two [realizing] states
to be a single 'physical-chemical state' . . . is not a metaphysical option
that can be taken seriously" (Rosenthal, 1991, 201). I think Putnam is right
that it's not. But the question why it isn't remains.
Here's where we've got so far: Functionalists agree with Kim that there
are no laws about disjunctive properties, whether the disjunctions are
metaphysically open like being in pain, or metaphysically closed like being
jade. But, according to functionalists, there can be a projectible property
that is coextensive with an open disjunction (e.g., being in pain); whereas,
in the case of a closed disjunction, there can't be. So, what, according to
the functionalist, accounts for this asymmetry? In one form or another,
that question keeps coming up. It needs an answer.
Here's a possible story: Suppose, for a moment, that metaphysically
open disjunctive properties are projectible after all. Still, there's a differ-
ence between the case where a disjunction appears only in bridge bicondi-
tionals, and the case where it is, as it were, independently certified because
it also occurs in laws at its own level. It might be quite reasonable to hold
that disjunctions are bona fide in bridge laws only when their projectiblity
is independently certified in this way. That would distinguish "real" bridge
laws from those containing formulas that are not only disjunctive, but
also gerrymandered; the latter being cases where, intuitively, all that the
disjuncts have in common is that they realize some higher level state.
That was, in effect, the view I took in an earlier paper (also called
"Special Sciences," Fodor, 197 4). The objection I voiced there was not to
the projection of disjunctions-open or such, but rather to a
theory of interlevel nomic relations that fails to distinguish real reductions
from gerrymanderings. "Type" physicalism is supposed, by general con-
sensus, to be stronger than "token" physicalism; stronger, that is, than the
mere claim that all mental states are necessarily physically instantiated,
however unhomogeneously. I suggested that the difference is that type
physicalism, but not token physicalism, requires bridge laws that really
Special Sciences 17
are laws, namely, ones that contain predicates that are independently
certi6ed as projectible because they are independently required to state
intralevel laws. By this "no gerrymandering" criterion, it is empirically
plausible that the multiple realizers of functional states are often not pro-
jectible, either in psychology or elsewhere.
I still think that's perfectly all right, as far as it goes. In fact, I think it's
quite attractive since it fends off the following dilemma. I suppose a func-
tionalist might wish to admit that there are nomologically necessary con-
straints on what sorts of things can be realizers of pains (or can openers or
whatever). That is, for each functional property P, there must be some dis-
junction of realizers (call it "# ") such that "all Ps are R 1, or R2, or R3 ...
#" is nomologically necessary'; and even if we don't (or can't) know
what # is, presumably God can (and does). But if "all Ps are R1, or R2,
or R3 . . . #" is nomologicaly necessary, it look likes P is reducible after
all since it looks like "all Ps are R1, or R2, or R3 ... #" is a {bridge) law;
presumably, a law is just a universal conditional that's true in all nomo-
logically possible worlds. But the argument for the autonomy of meta-
physically open MR states depends on their not being lawfully related to
their realizers. So it looks like that argument must be unsound.
The way out of this for functionalists is to require that bridge laws be
not just nomologically necesary but also that they be not gerrymandered,
as per above. What's wrong with "all Ps are RI, or R2 ... " isn't that
it lacks a nomologically necessary closure; "#" is one by assumption.
What's wrong is rather that the nomological closure is gerrymandered. That
is, the predicate "R1, or R2, or ... #"isn't "independently certi6ed." That
is, it doesn't occur in any proper ("single level") laws. Since "R1, or R2, or
... #"isn't independently certi6ed, "all Ps are RI, or R2, or ... #"isn't a
bridge law and (ceteris paribus) P isn't reducible after all. On this account,
the constraints on bridge laws are stronger than (in effect, they include) the
constraints on proper (single-level) laws. This difference is what underlies
the intuition that type materialism comes to more than just the claim that
it's nomologically necessary that every nonbasic property be physically
realized.
As I say, I still like that story; but I admit that there may be more to
be said. It's clearly Kim's intuition that there's something wrong with
multiply based kinds as candidates for projectibility whether or not they
are gerrymandered. That is, that you aren't allowed multiply based kinds
either in inter- or in intralevellaws. For reasons we've already discussed, I
don't accept Kim's account of why this prohibition holds; on my view, his
diagnosis depends on his failure to distinguish closed disjunctions from
open multiple realizations. But I'm prepared to take it seriously that
maybe disjunctions are, as such, bad candidates for projection. And as
we've just seen, even functionalists have to claim, at a minimum, that
18 Chapter 2
nondisjunctive functional states are ipso facto better candidates for pro-
jection than are their open disjunctive realizers; otherwise open disjune-
tive laws about the latter would compete with nondisjunctive laws about
the former. So, then, just what is wrong with projecting disjunctions?
It's not hard to see why it's so plausible that there can't be laws about
closed disjunctions. By assumption, if P is the closed disjunction F v G,
then it is metaphysically necessary that the properties a thing has qua P
are either properties it has qua F or properties it has qua G; and, of course,
this includes projectible properties inter alia. That's why, if being jade
really is a closed disjunctive property (if being jade is just being jadeite or
nephrite) then of course there are no laws about being jade "as such"; all
the jade laws are ipso facto either jadeite laws or nephrite laws.
But it's not equally obvious why (or whether) there can't be laws
about open disjunctions as such. In fact, I think there are depths here to be
plumbed.
One might, for starters, try denying that open formulas-ones that
contain ellipses essentially-succeed in expressing properties at all; in
which case, their failure to express projectible properties would not be
surprising. But I propose to assume without argument that a predicate like
"is Rl v R2 ... " does pick out a corresponding property, namely, the
property of being Rl v R2 .... It is, remember, common ground between
Kim and the functionalists that mental states are multiply realized. So both
are committed to some sentence of the form "pain is Rl v R2 ... "being
true, and to its not being equivalent to "pain is Rl v R2." It's hard to see
how all that could be so unless predicates with ellpises can refer.
To take for granted that (openly) disjunctive sentences (can) have truth
conditions is not, of course, to say what their truth conditions are. I haven't
got a semantics for such sentences to offer you; and trying to construct
one is not a task that I relish. Maybe they should be treated as true at
each world where any of the disjuncts is true, and as neither true nor false
(as noncommital, if you like) anywhere else. In what follows, I'll assume
some story of that sort; I doubt the details matter much for the polemical
purposes I have in mind.
Where we are now is: We're assuming that open disjunctions can
express bona fide properties, but that the properties that they express
are somehow intrinsically unfit for projection. This can really seem quite
puzzling. Presumably God can do anything He likes with the properties
He has at hand; so if there really are such properties as being R1 v R2 ... ,
why can't God make laws about them? Nonetheless, I think the intuition
that open disjunctions are at best bad candidates for laws is basically
sound. Here's why: Open laws suggest missed generalizations. For, to offer a
law of the form {Rl v R2 v )--+ Q is to invite the charge that one has
Special Sciences 19
failed correctly to identify the property in virtue of which the antecedent
of the law necessitates the consequent. Or, to say the same thing the
other way around: Someone who offers such a law undertakes a burden to
provide a positive reason that there isn't a higher level but nondisjunctive
property of things that are Rl v R2 . . . in virtue of which they bring it
about that Q.
But we still haven't got to the bottom. No doubt, if there is a higher
level property that subsumes all the states that satisfy an open dis-
junction, then we will want to formulate our laws in terms of it. But why
is than Not, surely, because we are prejudiced against disjunctions "as
such''7 Rather, I think, it's because formulas that express closed laws are
stronger than the corresponding open ones, and ceteris paribus, we want
to accept the strongest generalizations that our evidence confirms. Accept-
ing the strongest generalizations that one's evidence confirms is what
induction is about.
I'm suggesting that the intuition that open disjunctions invite higher
level laws isn't different in kind from the intuition that open lists invite
universal generalizations. Swans are white constrains more worlds (hence
supports stronger counterfactuals) than x is white if (x is swan a, v x is swan
b, v x is swan c ... ). So too, and for the same reason, pain causes avoidance
constrains more worlds than x causes avoidance if (x is neural state a, v x is
neural state b ... , v x is silicon state f ... , v x is Martian state g, v ... ). The
cost of the former generalization is reifying not just the property of being
swan a or being swan b or being swan c, etc. but also the more abstract,
higher-level property of being a swan. The cost of the latter generalization
is reifying not just the property of being in neural state a or being in neural
state b ... or being in silicon state f . .. or being in Martian state g, etc. but also
the more abstract, higher-level property of being in pain. Pretty clearly,
standard inductive practice is prepared to hypostasize in both cases. And
the success of standard inductive practice suggests pretty clearly that it is
right to do so. ,
It may be, according to this view of the matter, that there are laws
about open disjunctive properties after all; God can do whatever He likes,
as previously remarked. But at least it's apparent that we have general
methodological grounds for preferring a nondisjunctive higher-level
law to a corresponding low:er-level disjunctive law, all else being equal.
Induction is a kind of market prudence: Evidence is expensive, so we
should use what we've got to buy all the worlds that we can.
Here's where the discussion stands now:
Kim is wrong about what's wrong with jade. What's wrong with
jade is not that it's MR but that it's a (closed) disjunctive property,
and closed disjunctive properties are ipso facto not projectible.
20 Chapter 2
Kim is also wrong about the analogy between jade and pain.
According to functionalists, pain, but not jade, is MR; that is, pain
has an openly disjunctive realizer. It is to that extent okay for a func-
tionalist to say both that there aren't laws about jade and that that
there are laws about pain.
But a functionalist who says this still needs to explain why we
should (why we do) prefer higher-level nondisjunctive laws (pain
leads to avoidance) to lower-level laws with open disjunctions (states
that are Rl v R2 v ... lead to avoidance), all else being equal. Why are
we prepared to buy nondisjundive laws at the cost of reifying high-
level properties? My story is that this policy complies with a dictate
that inductive practice obeys quite generally: Prefer the strongest
claim compatible with the evidence, all else being equal. Quantification
over instances is one aspect of rational compliance with this injunc-
tion; rei6cation of high level kinds is another.
So, then, everything is 6ne and all the mysteries-except, of course,
the ones about induction itself-have dissolved? Not! (as I'm horri6ed to
hear they say in California). I think, in fad, that what's really bugging Kim
is indeed a metaphysical mystery about functionalism, and that the dis-
cussion we've been having so far hardly touches it. Let me try, in the
closing section, to articulate what I take to be the trouble.
Conclusion (Molto Mysterioso)
Kim remarks, at one point, that "when we think about making projections
over pain, very much the same worry should arise about their propriety
as did for jade. Consider, a possible law: 'Sharp pains . . . cause anxiety
reactions.' Suppose this generalization has been well con6rmed for humans.
Should we expect on that basis that it will hold also for Martians whose
psychology is implemented (we assume) by a vastly different mechanism?
Not if we accept ... that psychological regularities hold, to the extent that
they do, in virtue of the causal-nomological regularities at the physical
implementation level" (324).
Apparently, Kim wants to get from premises 4 and S to the conclusion 6.
4. Psychological regularities hold only in virtue of implementation
level regularities.
S. Martian pain is implemented by "vastly different" mechanisms
than ours.
6. We shouldn't expect "pain" to be projedible over a mixed pop-
ulation of us and Martians; that is, we shouldn't expect that Martian
pain will be like our pain in respects that are thus far unexamined.
Spedal Sciences 21
But, on second thought, an enthymeme would seem to have crept into
this line of argument. To warrant his inference, Kim also needs some such
premise as 7.
7. The behaviors of systems that are "vastly different" at the physi-
cal level should be expected not to be reliably similar in respect of
their nonphysical (e.g., of their higher-level or functional) properties.
Such similarities as one finds are accidents (cf. the color of jadeA =
the color of jadeB).
But, in respect to (7), one has serious reason to hesitate. For one thing, arti-
facts appear to offer an indefinite number and variety of counterexamples;
that's why references to can openers, mousetraps, camshafts, calculators,
and the like bestrew the pages of functionalist philosophy of science. To
make a better mousetrap is to devise a new kind of mechanism whose be-
havior is reliable with respect to the high-level regularity "live mouse in,
dead mouse out." Ceteris paribus, a beHer mousetrap is a mechanism that
is even more reliable with respect to this regularity than mousetraps used
to be. If it weren't possible, at least sometimes, for quite heterogeneous
mechanisms to be reliable in respect of the functional descriptions that they
converge upon, new kinds of mousetraps would never work; ingenuity
would fail, and mousetrap makers would be doomed to careers of fruitless
self-quotation. It looks, in short, like (7) might do as a rule of thumb, but it
can't be anything like a natural necessity.
Likewise outside the realm of artifacts. The very existence of the special
sciences testifies to reliable macrolevel regularities that are realized by
mechanisms whose physical substance is quite typically heterogeneous.
Does anybody really doubt that mountains are made of all sorts of stuff?
Does anybody really think that, since they are, generalizations about
mountains as such won't continue to serve geology in good stead?
Damn near everything we know about the world suggests that unimagin-
ably complicated to-ings and fro-ings of bits and pieces at the extreme
microlevel manage somehow to converge on stable macrolevel properties.
On the other hand, this "somehow" really is entirely mysterious, and
my guess is that that is what is bugging Kim. He just doesn't see why
there should be (how there could be) macrolevel regularities at all
in a world where, by common consent, macrolevel stabilities have to
supervene on a buzzing, blooming confusion of microlevel interactions.
Or rather, he doesn't see why there should be (how there could be)
unless, at a minimum, macrolevel kinds are homogeneous in respect to
their microlevel constitution.
5
Which, however, functionalists in psychol-
ogy, biology, geology, and elsewhere, keep claiming is typically not the
case.
22 Chapter 2
So, then, why is there anything except physics? That, I think, is what is
really bugging Kim.
6
Well, I admit that I don't know why. I don't even
know how to think about why. I expect to figure out why there is any-
thing except physics the day before I figure out why there is anything at
all, another (and, presumably, related) metaphysical conundrum that I find
perplexing. I admit, too, that it's embarassing for a professional philos-
opher-a paid up member of the AP A, Eastern Division, I assure you-
to know as little as I do about why there is macrostructural regularity
instead of just physical regularity. I would therefore quite like to take
Kim's way out and dissolve the mystery by denying the premise. Kim
wants to just stipulate that the only kinds there are (what he calls) local;
that is, the only kinds there are the kinds of kinds whose realizers are
physically homogeneous. "[T]he present view doesn't take away species-
restricted mental properties ... although it takes away pain 'as such'" (25).
More precisely, it doesn't take away species-specific mental properties
unless it turns out that they are MR too; if they are, then of course it
does.
In effect, Kim wants to make it true by fiat that the only projectible
kinds are physically homogeneous ones. That's tempting, to be sure; for
then the question why there are macro level regularities gets exactly the
same answer as the question why there are microlevel regularities: Both
follow from physical laws, where a physical law is, by definition, a law that
has physical kinds on both ends. But, for better or worse, you don't get to
decide this sort of thing by fiat; just as you don't get to avoid the puzzle
about why there's something instead of nothing by stipulating that there
isn't. Only God gets to decide whether there is anything, and, likewise,
only God gets to decide whether there are laws about pains; or whether,
if there are, the pains that the laws are about are MR. Kim's picture seems
to be of the philosopher impartially weighing the rival claims of empirical
generality and ontological transparency, and serenely deciding in favor of
the latter. But that picture won't do. Here, for once, metaphysics actually
matters,
7
so philosophers don't get to choose.
Science postulates the kinds that it needs in order to formulate the
most powerful generalizations that its evidence will support. If you want
to attack the kinds, you have to attack the generalizations. If you want to
attack the generalizations, you have to attack the evidence that confirms
them. If you want to attack the evidence that confirms them, you have to
show that the predictions that the generalizations entail don't come out
true. If you want to show that the predictions that the generalizations
entail don't come out true, you have actually to do the science. Merely
complaining that the generalizations that the evidence supports imply a
philosophically inconvenient taxonomy of kinds cuts no ice at all. So far,
Special Sciences 23
anyhow, when the guys in the laboratories actually do the science, they
keep finding that mental kinds are typically MR, but that the predictions
that intentional psychology entails are, all the same, quite frequently con-
firmed. Lots of different sorts of microinteradions manage, somehow or
other, to converge on much the same macrostabilities. The world, it
seems, runs in parallel, at many levels of description. You may find that
perplexing; you certainly aren't obliged to like it. But I do think we had all
better learn to live with it.
Acknowledgment
Many thanks to Ned Block, Ernie Lepore, Joe Levine, Howard Stein, and
especially Barry Loewer for helpful criticism of an earlier draft.
Notes
1. From here on, I'll honor the distinction between psychological (etc.) states and psycho-
logical (etc.) terms only where it matters. For much of the discussion we're about to have,
it doesn't. Rou8hJy, psychological states are what the terms in psychological theories
denote if the theories are true.
l. "Real" clearly can't mean exceptionless in this context. But what Kim has against psycho-
logical laws isn't their failure to be strict. (Here again, contrast Davidson, 1980; see also
Fodor, 1991 and Schiffer, 1991.)
3. More precisely: If jade is nephrite or jadeite, then it's metaphysically necessary that jade
is nephrite or jadeite. I don't want to speculate on what w ~ would (should) do if, for
example, we were to find that there is yet a third kind of stuff in our jade samples.
4. I'm neutral on the hard question of whether a property can have a metaphysically possible
realizer that isn't one of its nomologically possible realizers.
S. Note that what's mysterious isn't macrostructure per se; it's irreducible (autonomous)
macrostructure. When macrokinds are metaphysically identical to microkinds, laws about
the latter imply laws about the former; likewise when macroregularities are logical or
mathematical constructions out of microregularities, as in the "Game of Ufe" described in
Dennett, 1991. Pace Dennett, such cases do not illuminate (what functionalists take to be)
the metaphysical situation in the special sciences. To repeat: autonomy implies "real" (viz.,
projectible) patterns without reduction.
6. To be sure, Kim is also bugged about problems of causal and explanatory over-
determination (see Kim, 1993, passim). But these are plausibly just the other side of the
metaphysical problem about levels. Microlevel properties are projectible by consensus. If
macrolevel properties can be both projectible and autonomous, it looks like a given causal
transaction could be irreducibly "covered" by more than one causal law; and, presumably,
reference to any of these covering laws could constitute a causal explanation of the trans-
action. That, I suppose, is what the issues about causal and explanatory overdetermination
amount to.
7. And not j\,\st for psychology; the parallels in the current evolutionary wars are positively
eerie. '1T]he ultra-Darwinians reveal a thoroughgoing reductionst stance. [They] simply
wave at large-scale systems, but only address the dynamics of gene-&equency changes as
they see them, arising from competitive reproductive struggle .... Naturalists, in contrast,
are attuned to the hierarchical structure of biological systems. They are convinced that
24 Chapter 2
there are processes relevant to undentanding evolution that go on within each of these
levels-&om genes, right on up through populations, species, and ecosystems" (5). 'The
implacable stablity of species in the face of all that genetic ferment is a marvelous demon-
stration that large-scale systems exhibit behavion that do not mirror exadly the events
and processes taking place among their parts .. Events and processes ading at any one
level cannot possibly explain all phenomena at higher levels (175). (Both quotes are &om
Eldredge, 1995. A super book, by the way, which everybody ought to reacl.)
Part II
Concepts
Chapter 3
Review of Christopher Peacocke's A Study of
'oncepts
The Modern era, as analytic philosophers reckon, started with Descartes.
By contrast, the Recent era started when philosophy took the '1inguistic
turn" (Richard Rorty' s phrase), hence with Frege or Russell, or early Witt-
genstein, or the Vienna Circle-take your pick. Modern philosophy was
mostly about epistemology; it wanted to understand what makes knowl-
edge possible. Recent philosophy is mostly about meaning (or "content")
and wants to understand what makes thought and language possible. So,
anyhow, we tell our undergraduates when we're in a hurry.
There's something to it, but probably not much. 'Transcendental" argu-
ments used to run: "If it weren't that P, we couldn't know that Q; and we
do know that Q; therefore P." Philosophical fashion now prefers: '1f it
weren't that P, we couldn't say (or think or judge) that Q; but we do say
(or think or judge) that Q; therefore P." Much of a muchness, really. The
two kinds of arguments tend to be about equally unconvincing and for
the same reasons. Often enough, empiricist preconceptions are doing the
work in both.
This is not, however, to deny that there is something very peculiar
about Recent philosophy. There has indeed been a change, and it goes
much deeper than shifting styles of philosophical analysis. What's really
happened, not just in philosophy but in psychology, lexicography, lin-
guistics, arti6cial intelligence, literary theory, and just about everywhere
else where meaning and content are the names of the game, is a new
consensus about what concepts are. Take a sample of current and Recent
theorists, chosen with an eye to having as little else in common as may
be: Heidegger, or Wittgenstein, or Chomsky, or Piaget, or Saussure, or
Dewey, or any cognitive scientist you like, to say nothing of such con-
temporary philosophers as Davidson, Dennett, Rorty, and Quine. You
may choose practically at random, but they are all likely to agree on this:
concepts are capacities; in particular, concepts are epistemic capacities.
Christopher Peacocke's A Study of Concepts is about as subtle and sophis-
ticated an elaboration of the idea that concepts are epistemic capac-
ities as you will ever want to read. It may, in fact, be a more subtle and
28 Chapter 3
sophisticated elaboration than you will ever want to read. Peacocke is
hard work and he spares his reader nothing. His prose is not, perhaps,
denser than the intricacy of his thought requires, so I'm warning, not
complaining; but his book wants exegesis, and it will surely get a lot.
Many's the graduate seminar that will slog its way through, line by line,
and will be edified by doing so.
I won't attempt anything of that sort here. There are too many pas-
sages that I do not understand, and of the ones I do understand, there are
too many that I haven't made up my mind about. It does seem to me,
however, that a striking number of Peacocke's moves depend upon
assumptions that he makes, explicitly but without argument, in
the book's first several pages. I propose to concentrate on these since, as
I remarked, they strike me as a symptom of our times.
Peacocke's topic is the nature of concepts. Just roughly and by way of
orientation:
1. Concepts are word meanings.
1
The concept DOG is what the word
"dog" and its synonyms and translations express. This ties theories
of concepts to theories of language.
2. Concepts are constituents of thoughts. To think that dogs bark is to
entertain the concept DOG and the concept BARK.
3. Concepts apply to things in the world. The concept DOG is one
which, of necessity, all and only dogs fall under. Judgments are
applications of concepts, which is why it's things in the world that
make judgments true or false.
This catalogue is mine, not Peacocke's, but I don't expect it's anything
that he'd object to very much. So then, if that's what concepts are, what
should a theory of concepts be7
Starting on page 5: ''Throughout this book I will try to respect the fol-
lowing principle .... There can be nothing more to the nature of a concept
than is determined by ... a correct account of 'grasping the concept' ....
[A] theory of concepts should be a theory of concept possession." There
are, to be sure, trivializing readings of this equation ( C is the unique con-
cept whose possession condition is that you have the concept C). But
Peacocke intends that the nature of a concept should be illuminated by
what a theory says about grasping it. For example (6), "Conjunction is
that concept C to possess which a thinker must find [inferences of certain
specified forms] primitively compelling, and must do so because they are
of these forms." For example, it partially identifies C as the concept of
conjunction that anybody who has it finds inferences from the premises p
and q to the conclusion pCq primitively compelling as such.
Peacocke is saying that sometimes part of what grasping a concept
comes to is being able to see, straight off, that certain of the inferences
Review of Peacocke's A Study of Concepts 29
that it figures in are okay. Another part of what grasping a concept some-
times comes to is being able to see straight off that the concept applies
to something one perceives. (This figures largely in the book, but I'll scant
it here. See the next two chapters in this volume.)
In either case, though Peacocke's main topic is the nature of conceptual
and/or linguistic content and is thus nominally about semantics, it's epis-
temology that is actually calling the shots. According to Peacocke, what
concepts you have depends on what concepts you have grasped, and
what concepts you have grasped depends on what epistemic capacities you
have acquired. Having the concept of CONJUNCTION, for example, is
being able to see certain inferences as valid. My point is that the putative lin-
guistic tum, from a kind of philosophy that worries mostly about knowl-
edge to a kind of philosophy that worries mostly about meaning, doesn't
actually amount to much if meaning is itself epistemically construed.
These days, the ideas that theories of concepts are theories of concept
possession and that possessing a concept is having certain epistemic
capacities are often treated as truisms. In fact, they are intensely tenden-
tious. I suppose Descartes, or Hume, or Mill would have thought that you
identify a concept not by saying what it is to grasp it but by saying what it
is the concept of. Accordingly, on this older view, to illuminate the nature
of a concept you need not a theory of concept possession but a theory of
representation. The key question about, as it might be, the concept DOG
is not "what is it to have it," but something like '1n virtue of what does
that concept represent dogs, and in virtue of what do other concepts fail
to do sor'
Changing the topic from "How do concepts represent?" to "What
capacities constitute the possession of a concept?" was, I think, what really
started Recent philosophy. And arguably Dewey and the Pragmatists had
more to do with it than Frege or Wittgenstein. It's a paradigmatically
Pragmatist idea that having a concept consists in being able to do some-
thing. By contrast, uninstructed intuition suggests that having a concept
consist in being able to think something. (Having the concept DOG is
being able to think about dogs; or, better, about the property of being a
dog.) In my view, it's uninstructed intuition that has this stick by the right
end.
So that joins the issue: Is a theory of concepts a theory of concept pos-
session or is it a theory of how concepts represent?"Why not both?" you
might ask, in one of your ecumenical moods. But this would be to miss
the metaphysical tone of Peacocke's inquiry. No doubt there might be
both a theory of what it is to grasp a concept and a theory of how a con-
cept represents. And there might also be theories of how a concept is
acquired and how it's applied, while we're at it. But, as Peacocke under-
stands things, none of these would be a conceptual analysis-none of
30 Chapter 3
them would count as philosophy, strictly speaking-unless it specifies the
properties that make a concept the very concept that it is. Peacocke is
saying that what makes a concept the very concept that it is are the con-
ditions for possessing it.
Because he thinks philosophical analyses unpack the possession con-
ditions that individuate concepts, Peacocke is prepared to take quite a
hard line on the methodological priority of philosophical investigations in
the cognitive sciences.
An agenda for psychology suggested by the general approach
I have been advocating is, then, this: For each type of thinker and
for each concept possessed by a thinker of that type, to provide a
subpersonal explanation of why the thinker meets the possession
condition for that concept .... Carrying out this agenda is also in its
very nature an interdisciplinary enterprise. For any particular con-
cept, the task for the psychologist is not fully formulated until the
philosopher has supplied an adequate possession condition for it.
(190)
The idea that philosophy sets the agenda for psychology, or for any
other empirical inquiry, is a typical product of the idea that philosophy is
conceptual analysis and that conceptual analysis is, as Peacocke likes to
say, "relatively a priori." It strikes me, frankly, as ahistorical and maybe a
touch hubristic. The susurration that you hear is .legions of cognitive psy-
chologists not holding their breath until their task is fully formulated by
philosophers.
I doubt that the theory of concepts that engenders this account of con-
ceptual analysis can be sustained. The key problem is that people who
have the concept DOG thereby have all sorts of capacities that people
who don't have that concept thereby fail to have. And, surely, not all
these capacities are essential to "the nature of the concept." If I didn't
have the concept DOG, I suppose I couldn't have the concepts DOG
BATH or DOG BONE or FIGHTING LIKE CATS AND DOGS. And, if I
didn't have those concepts, there would be all sorts of inferences that I
would fail to find "primitively compelling" (from "dog bath" to "bath for
dogs," for example) and all sorts of perceptual judgments that I would be
unable to make (that the object .currently on display is a dog bath rather
than a bird bath, for example). But I suppose (and so, I'm sure, would Pea-
cocke) that none of these capacities illuminates the essential nature of the
concept DOG; one could have the concept even if one had none of them.
If there is a capacity the possession of which is constitutive of grasping
the concept DOG, and if having this capacity consists, inter alia, in being
able to see certain inferences as primitively compelling, then there must
Review of Peacocke's A Study of Concepts 31
be something that distinguishes such concept-constitutive inferences from
the rest. And, as we've seen, it has to be something nontrivial if con-
ceptual analysis is to be worth doing. (It's constitutive but trivial that
DOG is a concept that lets you make inferences about dogs. It's non-
trivial, but also nonconstitutive, that DOG is a concept that lets you make
inferences about dog baths.) Are there some inferences that are both non-
trivial and constitutive of the concept DOG? The answer must be "yes" if
theories of concepts are theories of their possession conditions and pos-
session conditions are inferential capacities.
Very well then, what determines which inferential capacities are non-
trivially constitutive of grasping the concept C7 Which of the cluster of
capacities that grasping C may bring in train are the ones that belong to
its possession conditions? Peacocke says a lot about which possession
conditions are constitutive of one or other concepts, but remarkably little
about the general question.
The closest we get is this (2): "Concepts C and D are distinct if and
only if there are two complete propositional contents that differ at most
in that one contains C substituted in one or more places for D, and one of
which is potentially informative while the other is not." So, for example,
the concept DOG is distinct from the concept BARKER because someone
who has fully grasped the former concept, and who takes it that dogs are
animals, might nevertheless take it to be news that barkers are animals.
Whereas (assuming that DOG is the same concept as DOMESTIC
CANINE) nobody who takes it that dogs are animals could find it news
that domestic canines are too. Or, to put the same idea in terms of pos-
session conditions: finding the inference if dogs are animals then domestic
canines are animals primitively compelling is among the possession con-
ditions for DOG. Whereas, though the inference if dogs are animals then
barkers are animals is quite a good inference, finding it primitively com-
pelling would presumably not count as constitutive for any of the con-
cepts involved. This would be true even if, in point of fact, all and only
dogs bark. (I emphasize that the example is mine and not Peacocke's, and
that it is crude.)
The notion of an informative proposition (or inference) thus looms
very large in Peacocke's treatment. He needs it a lot if he is to avoid the
trivialization of his project. As far as I can tell, Peacocke thinks it's too
obvious to bother arguing for that you can. individuate possession con-
ditions, and thereby flesh out the notion of concept identity, by appealing
to the informativeness test. But I don't at all share his optimism.
Someone who finds it unsurprising that John understands that bachelors
are bachelors might, I suppose, still wonder whether John understands
that bachelors are unmarried men. So it appears that if, following Pea-
cocke's recipe, you substitute "unmarried men" for the second "bachelors"
32 Chapter 3
in '1ohn understands that bachelors are bachelors" you go from some-
thing unsurprising to something that someone might well take to be news.
Since, however, the concepts BACHELOR and UNMARRIED MAN are
identical if any concepts are, it looks like the informativeness test for con-
cept identity is badly undermined. These so-called Mates cases (after their
inventor, Benson Mates) are a philosophical commonplace; Peacocke doesn't
mention them, but I don't understand why they don't worry him.
Or consider poor Jones, who went off with a bang; he didn't know that
being flammable and being inflammable are the same thing. Jones would
have found it informative had someone taken the trouble to tell him; if
someone had, he would be with us still. Yet "flammable" and "inflamma-
ble" are synonyms and hence must express the same concept if concepts
are word meanings. It looks, again, as though informativeness is one
thing, conceptual identity another. (Peacocke has a remark on p. 32,
arising in a quite different context, that may be intended to cover this
sort of case: '1n this particular example, it suffices for a theory of concepts
to aim to explain those patterns of epistemic possibility that exist only for
one who fully understands [the corresponding word] (and any synonyms he
may acquire)" [Peacocke's parentheses, my emphasis]. I doubt, however,
that Peacocke intends this as a codicil to his informativeness test, since it
presupposes a notion of synonymy which is itself semantic and, to put it
mildly, unexplicated.)
I am, truly, not meaning to quibble or to insist upon what are arguably
marginal counterexamples. But, like lots of other philosophers who have
been influenced by Quine, I really do doubt that concept identity can be
explicated un-question-beggingly by appeal to notions like informative-
ness. I doubt, in fact, that it can be un-question-beggingly explicated at
all so long as you think of having a concept in terms of possessing diag-
nostic epistemic capacities. (Whether concept identity can be explicated
un-question-beggingly in nonepistemic terms is a long question to which
the short answer is "maybe.") My own guess is that there aren't any non-
trivial inferences that the concept DOG requires its possessors to have
come what may. As Hilary Putnam has pointed out, even dogs are animals
would fail in science-fiction worlds where dogs tum out to be Martian
robots. In such a world, somebody could "fully grasp" the concept DOG
but find the inference dog ~ animal uncompelling; indeed, unsound. Nor, I
think, are there any perceptual judgments that DOG owners as such must
be compelled to make. No landscape is so uncluttered that it is impossible
in principle that one should fail to recognize that it contains a dog.
The problem is that Peacocke's whole project, his whole conception of
what concepts are, and hence of what a theory of concepts should aim for,
is committed to an epistemic distinction between analytic (constitutive)
Review of Peacocke's A Study of Concepts 33
capacities and synthetic (collateral) capacities. And this is a distinction that
Peacocke, like the rest of us, doesn't know how to draw. (He takes note
of this commitment in a passing footnote on p. 243 and is unperturbed
by it.) This leads to a geographical impasse: If, as people on my side of
the Atlantic are increasingly inclined to suppose, there isn't an epistemic
analytic/synthetic distinction, then the notion of a possession condition is
infirm and you can't identify grasping a concept with being disposed to
draw the inferences by which its possession conditions are constituted.
Correspondingly, the philosophical analysis of a concept can't consist in
setting out the possession conditions of the concept.
The long and short is: I think there is good reason to doubt that the
kind of philosophy Peacocke wants to do can be done. In one passage,
Peacocke remarks almost plaintively that "theories are developing in the
literature of what it is to possess certain specific concepts .... While there
is much that is still not understood and not all of what has been said is
right, it is hard to accept that the goal of this work is completely mis-
conceived" (35-:-36). I guess I don't find it all that hard. The linguistic tum
was, I think, an uncompleted revolution; to really tum from theories of
knowledge to theories of meaning, you would have to stop construing
content in epistemological terms. Many analytic philosophers can't bear
not to construe content in epistemological terms because they think of
philosophy as conceptual analysis, and of conceptual analysis as display-
ing a concept's possession conditions, and of possession conditions as
characteristically epistemic.
2
If, as I believe, that whole picture is wrong, a
certain kind of analytic philosophy is ripe for going out of business. If
there is no analytic/synthetic distinction, then there are no analyses. This is a
thought that keeps philosophers on my side of the Atlantic awake at
night. Why doesn't it worry more philosophers on Peacocke's side7
Does any of this really matter except to philosophers over sherry?
Oddly enough, I think perhaps it does. We are in the midst of a major
interdisciplinary attempt to understand the mental process by which
human behavior accommodates to the world's demands-an attempt to
understand human rationality, in short. Concepts are the pivot that this
project turns on since they are what mediate between the mind and the
world. Concepts connect with the world by representing it, and they con-
nect with the mind by being the constituents of beliefs. If you get it
wrong about what concepts are, almost certainly you will get the rest
wrong too.
The cognitive scientists I know are mostly a rowdy and irreverent lot,
and I shouldn't want to be around when they hear Peacocke's views about
the primacy of philosophy in defining their enterprise. But it's perfectly
true that they have, almost without exception, assumed what is essentially
a philosophical theory of concepts, and that it's pretty much the one that
34 Q:tapter 3
Peacocke also takes for granted: concepts are epistemic capacities. In con-
sequence, questions about which epistemic capacities constitute which
concepts perplex the whole discipline, and nobody knows any more than
Peacocke does how to answer them. If it turns out that concepts aren't
epistemic capacities, these questions don't have answers.
I'm not proposing a trans-Atlantic methodological shoot-out, but I do
think there needs to be a sustained discussion of what concepts are, and
I think that de-espistemologizing semantics-completing the linguistic
tum-is likely to be its outcome. If so, theories of language and mind will
eventually come to look very different from what Recent philosophy has
supposed; and the project of philosophical analysis will come to look
inconceivably different. (Assuming, indeed, that there is any such project
left.) In this discussion, someone will have to speak with insight and
authority for the Received View. Peacocke has done that, and we are all in
his debt.
Notes
1. Which is not to say that word meanings are ipso facto concepts; I suppose Peacocke
would deny that they are in the case of demonstratives or names, for example. Nor does
he claim (nor should he) that for each of our concepts we must have a corresponding
word with which to express it.
Getting clear on the word-concept relation is no small matter. Suffice it, for purposes of
the present discussion, that Peacocke thinks that at least some words mean what they do
because they express the concepts that they do; and I think he's right.
2. It's not, of course, unintelligible that analytic philosophers should have asswned all this.
Transcendental arguments are supposed to ground antiskeptical conclusions. The plan is
to say to the skeptic something like the following: "Look. if you don't recognize even this
as a dog (or, mutatis mutandis, "if you don't accept even that dogs are animals'') then you
simply haven't got the concept. I can only argue about dogs with someone who does have
the concept." But, of course, that kind of line won't work unless the connection between
the concept and the corresponding epistemic capadties is constitutive.
Lots of philosophers fear that if concepts don't have analyses, justification breaks down.
My own guess is that concepts don't have analyses and that justification will survive all
the same.
Chapter 4
There Are No Recognitional Concepts-Not Even
RED
Introduction
Let it be that a concept is recognitional if and only if:
1. It is at least partially constituted by its possession conditions; and
2. Among its possession conditions is the ability to recognize at
least some things that fall under the concept as things that fall under
the concept.
For example, RED is a recognitional concept iff it numbers, among its
constitutive possession conditions, the ability to recognize at least some
red things as red.
In this paper, I propose to argue-indeed, I propose to sort of prove-
that there are no recognitional concepts; not even RED.
Lots of philosophers are sympathetic to the claim that there are recog-
nitional concepts. For one thing, insofar as recognitional capacities are
construed as perceptual capacities, the claim that there are recognitional
concepts preserves the basic idea of Empiricism: that the content of at
least some concepts is constituted, at least in part, by their connections
to percepts. For philosophers who suppose that Empiricism can't be all
wrong, recognitional concepts can therefore seem quite a good place to
dig in the heels. More generally, the claim that there are recognitional
concepts is a bastion of last resort for philosophers who think that
semantic facts are constituted by epistemological facts, a doctrine that
includes, but is not exhausted by, the various forms of Empiricism. If you
have ever, even in the privacy of your own home among consenting
adults, whispered, hopefully, the word "criterion," then probably even you
think there a r ~ recognitional concepts. .
Philosophers who hold that there are recognitional concepts generally
hold that it's important that there are; for example, a familiar line of anti-
skeptical argument turns on there being some. The idea is that, if a con-
cept is recognitional, then having certain kinds of experience would, in
principle, show with the force of conceptual necessity that the concept applies.
If, for example, RED is a recognitional concept, then having certain kinds
36 Chapter 4
of experience would, in principle, show with the force of conceptual
necessity that there are red things. Ditto, mutatis mutandis, SQUARE,
CHAIR, IS IN PAIN, and BELIEVES THAT P, assuming that these are
recognitional concepts. So, if you think. that it's important that skepticism
about squares, chairs, pains, beliefs, or red things be refuted, you are likely
to want it a lot that the corresponding concepts are recognitional. Never-
theless, it's sort of provable there aren't any recognitional concepts; so, at
least, it seems to me.
I pause to mention a kind of argument against there being recognitional
concepts to which I am sympathetic, but which I am not going to pursue
in what follows: namely, that it's truistic that the content of one's experi-
ence underdetermines the content of one's beliefs, excepting only one's
beliefs about one's experiences. No landscape is so empty, or so well lit-
so the thought goes-that your failure to recognize that it contains a
rabbit entails that you haven't got the concept RABBIT. So, it couldn't be
that your having the concept RABBIT requires that there are circum-
stances in which you couldn't but recognize a rabbit as such.
I think this is a good argument, but, notoriously, lots of philosophers
don't agree; they think, perhaps, that the connection between concept
possession and recognitional capacities can be relaxed enough to accom-
modate the truisms about rabbits without the claim that there are recog-
nitional concepts lapsing into vacuity. I propose, in any event, not to rely
upon this sort of argument here.
Compositionality
The considerations I will appeal to are actually quite robust, so a mini-
mum of apparatus is required to introduce them. It will, however, be use-
ful to have on hand the notion of a satisfier for a concept. The satisfier(s)
for a concept are the states, capacities, dispositions, etc. in virtue of which
one meets the possession condition(s) for the concept.
1
So, if the ability to
tell red from green is a possession condition for the concept RED, then
being able to tell red from green is a satisfier for the concept RED. If a dis-
position to infer P from P & Q is a possession condition for the concept
conjunction, then being so disposed is a satisfier for the concept conjunc-
tion. And so forth. Since, by assumption, concepts have their possession
conditions essentially, and possession conditions have their satisfiers essen-
tially, the exposition will move back and forth between the three as
convenience dictates.
I propose to argue that there are no concepts among whose satisfiers
are recognitional capacities, hence that there are no recognitonal concepts.
I need a premise. Here's one:
There Are No Recognitional Concepts 37
Premise P: 5 is a satis6er for concept C if and only if C inherits 5
from the satis6ers for its constituent concepts.
1
(I'll sometimes call
this the "compositionality condition" on concept constitution.)
Consonant with my general intention not to have the argument tum on
its details, I leave it open how "inherited from" is construed in premise P,
so long as Axing the satis6ers for constituent concepts is necessary and
sufficient for Axing the satis6ers for their hosts. I do, however, urgently
call your attention to the following point, ignoring which breeds mon-
sters. Suppose concept Cis a constituent of concept H(ost). Suppose, also,
that S is a satis6er for C. Then, of course, perfectly trivially, since you
can't have H unless you have C, you can't have H unless you have a con-
cept one of whose satis6ers is S. What does not follow, however, is that
you can't have H unless you have S. To get that conclusion you need
the further principle that you can't have a concept unless you have the
satis6ers for its constituents; that is, the principle that hosts "inherit"
the satis6ers of their constituents. This is untrivial, and it is what premise
P asserts.
Why Premise P is Plausible
Unless P is true, we will have to give up the usual account of why con-
cepts are systematic and productive; and, mutatis mutandis, of how it is
possible to learn a language by learning its Anite basis. Consider, for
example, the concept-constitutive possession conditions for the concept
RED APPLE. If premise P is false, the following situation is possible: The
possession conditions for RED are ABC and the possession conditions for
RED APPLE are ABEFG. So denying P leaves it open that one could have
the concept RED APPLE and not have the concept RED.
But, now, the usual compositional account of productivity requires that
one satisfy the possession conditions for complex concepts, like RED
APPLE, by satisfying the possession conditions for their constituent con-
cepts. That is, it requires that one's having a grasp of the concept RED is
part of the explanation of one's having a grasp of the concept RED APPLE.
So accepting the usual compositional account of productivity is incom-
patible with denying premise P.
Likewise the other way around. The usual compositional account of
productivity requires that if one satis6es the possession conditions for the
constituents of a complex concept, one thereby satis6es the possession
conditions for the concept.
3
But, suppose that premise P is false, and con-
sider, once again, the concept RED APPLE. Denying P leaves it open that
the concept-constitutive possession conditions for RED APPLE are not
exhausted by the concept-constitutive possession conditions for RED and
APPLE. For example, the former might be ABCDE and the latter might
38 Chapter 4
be AB and CD respectively. But then grasping the concepts RED and
APPLE would not be sufficient for grasping the concept RED APPLE,
and, once again, the standard account of conceptual productivity would
be undermined.
So much for the bona fides of premise P. The next point is that the
condition that compositionality imposes on concept constitution is highly
substantive. A brief digression will show the kind of theory of concepts
that it can rule out.
Consider the idea that concepts are (or are partially) constituted by
their stereotypes, hence that knowing its stereotype is a satisfier for some
concepts. Premise P says that this idea is true only if, if you know the
stereotypes for the constituents of a complex concept, then you know
the stereotype for that concept. Which, in some cases, is plausible enough.
Good examples of RED APPLES are stereotypically red and stereotypi-
cally apples. Let's assume that that's because the stereotype of RED APPLE
is inherited from the stereotype for RED and the stereotype for APPLE. (This
assumption is concessive, and it may well not be true. But let it stand for
the sake of the argument.) So then, as far as RED APPLE is concerned, it's
compatible with premise P that knowing their stereotypes should be pos-
session conditions for RED and APPLE.
Still, concepts can't be constituted by their stereotypes (knowing its
stereotype can't be a satisfier for a concept). That's because RED APPLE
isn't the general case. In the general case, complex concepts don't inherit
their stereotypes from those of their constituents. So, in the general case,
stereotypes don't satisfy premise P.
Consider such concepts as PET FISH, MALE NURSE, and the like.
4
You
can't derive the PET FISH stereotype from the FISH stereotype and the
PET stereotype. So, if stereotypes were constitutive of the corresponding
concepts, having a grasp of FISH and having a grasp of PET (and know-
ing the semantics of the AN construction; see fn. 3) would not suffice for
having a grasp of PET FISH. So, the usual story about how PET FISH is
compositional would fail.
So much for stereotypes. If premise P is true, it follows that they can't
be concept-constitutive. I will now argue that, if premise Pis true, then it
likewise follows that there are no recognitional concepts.
In fact, most of the work is already done, since for all intents and pur-
poses, the notion of a recognitional concept is hostage to the notion that
concepts are constituted by their stereotypes. Here's why. Nobody could
(and nobody does) hold that the possession of a recognitional concept
requires being able to identify each of its instances as such; if that were the
requirement, then only God would have any recognitional concepts. So,
the doctrine must be (and, as a matter of fact, always is) that possession of
a recognitional concept requires the ability to identify good instances as
There Are No Recognitional Concepts 39
such in favorable conditions. (There are various variants of this in the
literature; but it doesn't matter to what follows which you choose.)
5
But now, unsurprisingly, the ability to recognize good instances of Fs
doesn't compose, and this is for exactly the same reason that knowing
the stereotype of F doesn't compose; good instances of F 8t Gs needn't be
either good instances ofF or good instances of G. See PET FISH, once
again: Good instances of PET FISH are, by and large, poorish instances of
PET and poorish instances of FISH. So a recognitional capacity for good
instances of PET and good instances of FISH is not required for, and typi-
cally does not provide, a recognitional capacity for good (or, indeed, any)
instances of PET FISH.
Somebody who is good at recognizing that trouts are fish and that
puppies are pets is not thereby good at recognizing that goldfish are pet fish.
The capacity for recognizing pet fish as such is not conceptually, or lin-
guistically, or semantically connected to capacities for recognizing pets as
such or fish as such. The connection is at best contingent, and it's entirely
possible for any of these recognitional capacities to be in place without
any of the others.
This doesn't, of course, show that the semantics of PET FISH are
uncompositional. What it shows is that recognitional capacities aren't
possession conditions for the concepts that have them. If recognitional
capacities were possession conditions, PET FISH would not inherit its
satisfiers from those of PET and FISH. So if recognitional capacities
were possession conditions, PET FISH would fail premise P. So recog-
nitional capacities aren't possession conditions. So there are no recogni-
tional. concepts.6
Objections
Q1: You're, in effect, taking for granted not only that compositionality
is needed to explain productivity, but that it is therefore a test for
whether a property is constitutive of the concepts that have it. Why
should I grant that?
A1: I suppose I could just dig my heels in here. Compositionality is
pretty nearly all that we know about the individuation of concepts. If we
give that up, we will have no way of distinguishing what constitutes a
concept from such of its merely contingent accretions as associations,
stereotypes, and the like.
But though I do think it would be justifiable to take that strong line, I
really don't need to in order to run the sort of argument I'm endorsing. If
push comes completely to shove, the following exiguous version will do
for my polemical purposes:
40 Chapter 4
If you know what "pet" and "fish" mean, you thereby know what
"pet fish" means. But you can be able to recognize pets and fish as
such but be quite unable to recognize pet fish as such. So recogni-
tional capacities can't be meanings, and they can't be constituents
of meanings, all varieties of Empiricist semantics to the contrary
notwithstanding.
As far as I can see, that formulation doesn't need much more than the
distinctness of discemibles, so it seems to me that it cuts pretty close to
the bone.
Q2: But couldn't an Empiricist just stipulate that recognitional capacities,
though they demonstrably don't satisfy premise P and are thus demon-
strably not constituents of meanings, are nevertheless to count as essen-
tial conditions for the possession of primitive concepts?
A2: Sure, go ahead, stipulate; and much joy may you have of it. But
nothing is left of the usual reasons for supposing that the concepts we
actually have comply with the stipulation; in particular, nothing is left of
the idea that the content of our concepts is constituted by our recogni-
tional capacities. Whatever content is, it's got to be compositional; so it's
got to come out that the content of RED APPLE includes the content of
RED and the content of PET FISH includes the content of PET.
It may be worthwhile to reiterate here an argument I gave for the plau-
sibility of premise P. Suppose a primitive concept has a possession con-
dition that is not inherited by one of its complex hosts; suppose, for
example, that being able to recognize good instances of pets is a pos-
session condition for PET but is not a possession condition for PET FISH.
Then presumably it is possible that someone who has the concept PET
FISH should nonetheless not have the concept PET. I take this to be a
reductio, and I think that you should too.
Here's a closely related way to make the same argument: Perhaps
you're the sort of philosopher who thinks it's a possession condition for
RED APPLE that one is prepared to accept the inference RED APPLE ~
RED (i.e., that one finds this inference "primitively compelling").
7
If so,
that's all the more reason for you to hold that the possession conditions
for RED APPLE must include the possession conditions for RED. Hence
it's all the more reason for you to hold that the satisfiers for RED are
inherited under composition by RED APPLE. But if that's right, then once
again it couldn't be that a recognitional capacity is a satisfier for RED
unless it's a satisfier for RED APPLE. But the capacity to recognize pets as
such is not a satisfier for the concept PET FISH, so it can't be a satisfier for
PET. Since sauce for the goose is sauce for the gander, the ability to rec-
ognize red things is likewise not a satisfier for RED.
There Are No Recognitional Concepts 41
Q3: Couldn't we split the difference? Couldn't we say that the satisfiers
for the primitive concepts include recognitional capacities, but that the
satisfiers for complex concepts don't?
A3: Simply not credible. After all, people who have the concept PET
FISH do generally have a corresponding recognitional capacity; for exam-
ple, they are generally good at recognizing goldfish as pet fish. And,
surely, being able to recognize (as it might be) a trout as a fish stands in
precisely the same relation to having the concept FISH as being able to
recognize a goldfish as a pet fish does to having the concept PET FISH.
So, how could that relation be constitutive of concept possession in the
one case but not in the other? Is it, perhaps, that the concepts FISH and
PET FISH have content in different senses of "content"?
This sort of point is probably worth stressing. Some philosophers have
a thing about recognitional capacities because they want to tie meaning
and justification close together (see the remarks earlier about antiskeptical
employments of the idea that there are recognitional concepts). But if
recognitional capacities are constitutive only of primitive concepts, then
the connection between meaning and justification fails in infinitely many
cases. It will thus be concept-constitutive that (ceteris paribus) it's trout-
looking is evidence for "it's a fish," but not concept constitutive that (ceteris
paribus) it's goldfishlooking is evidence for "it's a pet fish." What possible
epistemological use could such a notion of concept-constitutivity be put
to?
Q4: FISH and PET are only relatively primitive (they're only primitive
relative to PET FISH). What about absolutely primitive concepts like
RED? Surely the concept RED is recognitional even if neither FISH nor
PET FISH is.
A4: It's just more of the same. Consider RED HAIR, which, I will sup-
pose, is compositional (that is, not idiomatic) and applies to hair that is red
as hair goes. This view of its semantics explains why, though red hair is
arguably not literally red, still somebody who has RED and has HAIR and
who understands the semantic implications of the syntactic structure AN,
can figure out what "red hair'' means. So, prima facie, RED HAIR is com-
positional and the demands of productivity are satisfied according to the
present analysis. s
But notice, once again, that the productivity /compositionality of the
concepts does not imply the productivity /compositionality of the corre-
sponding recognitional capacities. Somebody who is able to say whether
something is a good instance of HAIR and whether something is a good
instance of RED is not thereby able to recognize a good instance of RED
HAIR. Well then, what does "red" contribute to the semantics of "red
hair''? Just what you'd suppose: it contributes a reference to the property
42 Chapter 4
of being red (as such). It's just that its doing that isn't tantamount to, and
doesn't entail, its contributing a recognitional capacity for (good instances
of) redness.
One's recognitional capacity for RED doesn't compose. So one's recog-
nitional capacity for red things is not a satis6er for the concept RED. So
not even RED is a recognitional concept.
Q5: What do you say about intentional concepts?
AS: Nothing much for present purposes. They have to be compositional,
because they are productive. If they are compositional, then there are, to
my knowledge, three theories (exhaustive but not exclusive) of what they
inherit from their constituents:
they inherit the extensions of their constituents;
they inherit the senses of their constituents;
the inherit the shapes of their constituents. (Notice that shape is
compositional; "red hair'' contains "red" as a morphosyntactic part;
and the shape of "red hair'' is completely determined given the
shape of "red," the shape of ''hair," and the morphosyntactics of the
expression.)
But, invariably a theory of intentional concepts that says that any of
these are inheritable properties of their constituents will also say that they
are constitutive properties of their constituents. So, as far as I can tell,
nothing that anybody is likely to want to say about intentional concepts
will deny my argument the premise it requires.
Conclusion
The moral of this chapter is that recognitional capacities are contingent
adjuncts to concept possession, much like knowledge of stereotypes; a
fortiori, they aren't constitutive of concept possession. How, indeed, could
anyone have supposed that recognitional capacities are satis6ers for con-
cepts, when recognitional capacities patently don't compose and concept
satis6ers patently do?
I think what went wrong is, after all, not very deep, though it's well
worth attending to. Content, concept-constitutivity, concept possession,
and the like, are connected to the notion of an instance (i.e., to the notion
of an extension). The notion of an instance (extension) is semantic, hence
compositional, through and through; idioms excepted, what is an instance
of a complex concept depends exhaustively on what are .the instances of
its parts. The notion of a recognitional capacity, by contrast, is connected to
the notion of a good (in the sense of a typical, or an epistemically reliable)
instance; the best that a recognitional capacity can promise is to identify
There Are No Recognitional Concepts 43
good instances in favorable conditions. It's a mistake to try to construe
the notion of an instance in terms of the notion of a good instance;
9
unsurprisingly, since the latter is patently a special case of the former, the
right order of exposition is the other way around.
Recognitional capacities don't act like satisfiers: What's a satisfier for a
complex concept depends on what's a satisfier for its parts; but what's
a good instance of a complex concept doesn't depend on what's a good
instance of its parts. Why should it? What's a good instance of a concept,
simple or complex, depends on how things are in the world.
1
Composi-
tionality can tell you that the instances of PET FISH are all and only the
pet fish; but it can't tell you that the good instances of pet fish are the
goldfish; which is, all the same, the information that pet fish recognition (as
opposed to mere PET FISH instantiation) is likely to depend on. How
could you expect semantics to know what kind of fish people keep for
pets? Likewise, what counts as red hair depends, not just on matters of
meaning, but also on what shades hair actually comes in (i.e., because red
hair is hair that is relatively red.) How could you expect semantics to know
what shades hair actually comes in? Do you think that semantics runs a
barber shop7
11
How, in short, could you expect that relations between
recognitional capacities would exhibit the compositionality that produc-
tivity requires of semantic relations?
_Qh, well; so what if there are no recognitional concepts?
For one thing, as I remarked at the outset, if there are no recognitional
concepts we lose a certain class of antiskeptical arguments; ones that
depend on the connection between percepts and concepts being, in some
sense, constitutive. We will no longer be able to say to the skeptic: "If
you don't think that this experience shows that that's a chair, then you
don't have the concept CHAIR." But maybe this isn't a great loss; I've
never heard of a skeptic actually being convinced by that kind of argu-
ment. I sure wouldn't be if I were a skeptic.
I'm not, however, meaning to deny that the issue about recognitional
concepts goes very deep. To the contrary, I'm meaning to claim that it
goes very much deeper than (mere) epistemology. Close to the heart of
the last hundred years of philosophy is an argument between a Cartesian
and a Pragmatist account of concept possession. Though the details vary,
the essentials don't: According to Cartesians, having the concept X is
being able to think about Xs; according to Pragmatists, its being able to
respond differentially or selectively to Xs (for short: it's being able to sort
Xs.) I doubt that there's a major philosopher, anyhow since Peirce-and
including, even, the likes of Heidegger-who hasn't practically every-
thing at stake on how this argument turns out.
Notice that the issue here isn't ''Naturalism." Sorting is just as inten-
tional as thinking, and in the same way: Neither coextensive thoughts
44 Chapter 4
nor coextensive sorts are ipso facto identical. A Pragmatist who isn't a
behaviorist can (and should) insist on this. The issue, rather, is whether
the intentionality of thought derives from the intentionality of action.
Roughly, Pragmatists think that it does, whereas Cartesians think that the
metaphysical dependencies go the other way around. It's the difference
between holding, on the one hand, that whether you are sorting Xs is a
matter of how you are thinking about what you are doing; or, on the
other hand, that whether you are thinking about Xs depends on (possibly
counterfactual) subjunctives about how you would sort them.
Well, the minimal Pragmatist doctrine (so it seems to me) is the claim
that there are recognitional concepts; that is, that at least some concepts
are constituted by one's ability to sort their instances. And the present
arguments (so it seems to me) show that even this minimal Pragmatist
doctrine isn't true. Thinking centers on the notion of an instance; recogni-
tional capacity centers on the notion of a good instance. Unless you are
God, whether you can recognize an instance of X depends on whether it's
a good instance of an X; the less good it is, the likelier you are to fail.
12
But you can always think an instance of X; namely, by thinking an
instance of X. So thinking is universal in a way that sorting is not. So
thinking doesn't reduce to sorting. That is bedrock. To try to wiggle
out of it, as so many philosophers have drearily done, by invoking ideal
sorts, recognition under ideal circumstances, the eventual consensus of the
scientific community, or the like, is tacitly to give up the defining Prag-
matist project of construing semantics epistemologically. Being an ideal sort
always turns out not to be independently definable; it's just being a sort
that gets the extension right.
Or, to put the point with even greater vehemence: The question
whether there are recognitional concepts is really the question what
thought is for; whether it's for directing action or for discerning truth.
And the answer is that Descartes was right: The goal of thought is to
understand the world, not to sort it. That, I think, is the deepest thing that
we know about the mind.
Afterword
This paper was presented at the 1997 meeting of the Central Division of
the American Philosophical Association. Stephen Schiffer commented, and
what he said was typical of the reaction I've had from a number of philo-
sophical friends. So I include here my reply to Steve's reply.
Steve asked, in effect: ''What's wrong with a mixed view, according to
which recognitional capacities are constitutive for (some) primitive con-
cepts but not for their complex hosts7' Steve thinks that my reply must
be either aesthetic (mixed theories are ugly) or an outright appeal to the
There Are No Recognitional Concepts 45
"agglomerative principle" that if the conjuncts of a conjunctive proposi-
tion are recognitional, then so too is the conjunctive proposition. Since
Steve takes this principle to be not better than dubious, he thinks that I
haven't a better than dubious argument against there being recognitional
concepts.
Now, it does seem to me that somebody who holds that there are
primitive recognitional concepts should also hold the agglomerative prin-
ciple (see the discussion of Q3). But my thinking this isn't an essential part
of my argument. I tried to make clear in the text what the essence of my
argument is; but, evidently, I didn't succeed. This hasn't been my century
for making things clear.
Here it is again:
A theory of compositionality should explain why, in the standard case,
anybody who has a complex concept also has its constituent concepts
(why anybody who has RED TRIANGLE has RED and TRIANGLE; why
anybody who has GREEN HAIR has GREEN and HAIR ... and so forth).
This is tantamount to saying that a compositionality principle should be
so formulated as to entail that satisfying the possession conditions for
a complex concept includes satisfying the possession conditions for its
constituents.
Now look at Steve's proposal, which is that "it is reasonable to hold
that the possession conditions of complex concepts are determined by
those of their constituents concepts. But for the case at hand this simply
requires F & G to be such that to possess it one must be able to recognize
good instances ofF and good instances of G."
As stated, Steve's theory is wrong about PET FISH:. It's true, of course,
that to have the concept PET FISH you have to have the concept FISH.
But it's certainly not true that to have the concept PET FISH you have to
have a recognitional capacity for good instances of fish. To have a con-
cept, conjunctive or otherwise, you have to have the concepts that are its
constituents. But you don't have to have recognitional capacities corre-
sponding to its constituents; not even if, by assumption, the complex concept
is itself recognitional. So, having the constituents of a concept can't require
having a recognitional capacity in respect of their instances. If it did, you
could have a complex concept without having its constituents-which is
not an option. So concepts can't be recognitional capacities.
Can Steve's proposal be patched? Well, if he is to get the facts to come
out right, he'll presumably just have to stipulate that for some F & G con-
cepts (RED TRIANGLE) "the possessor must be able to recognize good
instances of F and G," but that for others (PET FISH, MALE NURSE)
that's not required. And he'll have to say, in some general and principled
way, which concepts are which. But, surely, you don't want to have to
stipulate the relations between the possession conditions for a complex
46 Chapter 4
concept and the possession conditions for its constituents; what you want
is that they should just fall out of the theory of compositionality together
with the theory of concept constitutivity.
Which, indeeed, they do, if you get these theories right. What's con-
stitutive of FISH, and hence what PET FISH inherits from FISH, is (not a
capacity for recognizing fish but) the properly that FISH expresses: namely,
the properly of being a fish. Likewise, mutatis mutandis, what's constitutive
of RED, and hence what RED TRIANGLE inherits from RED, is (not a
recognitional capacity for red things but) the property that RED expresses,
namely, the property of being RED. Given that what is constitutive of a
concept determines its possession conditions, it follows that you can't
have the concept PET FISH unless you know that pet fish are fish, and
you can't have the concept RED TRIANGLE unless you know that red
triangles are red. This is, of course, just what intuition demands.
Explanations are better than stipulations; and they're a lot better than
stipulations that misdescribe the facts. So there still aren't any recogni-
tional concepts.
1
.3
Acknowledgment
Thanks to Ned Block, Paul Horwich, Chris Peacocke, Stephen Schiffer, and
Galen Strawson for helpful comments on an earlier draft of this essay.
Notes
1. Whereas, by contrast, the satis6ers of a concept are just whatever is in its extension. This
is not, admittedly, a very happy ytay of talking, but it's no worse, surely, than intention/
intension, and nothing better came to mind.
2. If C is a primitive concept, the condition is trivially satis6ed.
3. This isn't quite right, of course; you also have to know how the constituents are "put
together." Suppose a satis6er for RED is being able to identify red things and a satis6er
for SQUARE is being able to identify square things. Then, presumably, the correspond-
ing satis6er for RED SQUARE is being able to identify things in the intersection of RED
and SQUARE. The fad that it's the intersection rather than, say, the union, that's at issue
corresponds to the structural difference between the concept RED SQUARE and the
concept RED OR SQUARE.
4. The immediately following arguments are familiar from the cognitive science Uterature
on stereotypes, so I won't expand on them here. Suffice it to emphasize that the main
point-that stereotypes don't compose-holds whether stereotypes are thought of as
something Uke exemplars or as something Uke feature sets. For a review see Fodor and
Lepore, 1992.
S. The intended doctrine is that having the recognitional concept F requires being able to
recognize good instances of F as instances of F, not as good of F. It's concessive
of me to insist on this distinction, because it requires the empiricist to defend only the
weaker of the two views.
6. This assumes, of course, that what holds for PET and for FISH holds likewise for any
candidate recognitional concept: namely, that there will always be some complex concept
There Are No Recognitional Concepts 47
of which it is a constituent but to which it does not contribute its putative possession
condition. The reader who doubts this should try, as an exercise, to find a counter-
example. For starters, try it with RED and APPLE.
7. Even conceptual atomists like me can hold that inferences that relate a complex concept
to its parts are typically analytic and concept-constitutive. See Fodor and Lepore, 1992.
8. Let it be that an AN concept is "intersective" if its extension is the intersection of the As
with the Ns. The standard view is that being intersective is sufficient but not nececessary
for an AN concept to be compositional: RED HAIR is compositional but not intersectioe,
and PET FISH is both. (For a general discussion. see I<amp and Partee, 1995). Actually,
my guess is that RED HAIR. BIG ANT, and the like are the general case. Excepting the
"antifactive" adjectives ("fake," "imitation." etc.), AN usually means A FOR (an) N, and
the intersectives are just the limiting case where things that are A for (an) N are A.
But it doesn't matter for present purposes whether this is so.
9. Indeed, it's a venerable mistake. I suppose the Platonic theory of Forms was the 6rst to
commit it.
10. It also depends on how things are with us. What the good instances of RED are almost
certainly has to do with the way the physiology of our sensory systems is organized
(see, for example, Rosch, 1973; Berlin and Kay, 1969). Likewise, it's no accident that the
good instances of ANIMAL are all big enough for us to see (i.e., big enough for us to
see). It does not follow that a creature whose range of visual acuity is very different from
ours would thereby have a different concept of animals from ours.
11. The cases in the text are not, of course, exceptional. What counts as an average incOme
depends not only on what "average" and "income" mean. but also on what incomes
people actually earn. Semantics tells you that the average income is in the middle of the
income distribution. whatever the distribution may be. But if you want to recognize an
average income, you need the facts about how many people actually earn how much.
Semantics doesn't supply such facts; only the world can.
12. Analogous remarks hold for other epistemological capacities like, e.g., drawing infer-
ences. Unless you are God, whether in a particular case you are disposed to infer P from
P & Q depends, inter alia, on whether the logical form of the proposition is perspicuous.
This strongly suggests that the considerations that rule out recognitional capacities
as concept-constitutive will apply, mutatis mutandis, to rule out any epistemological
candidate.
13. Steve also suggested that maybe PAIN is a recognitional concept, even if RED is not. I
won't, however, discuss the notion that sensation concepts might be recognitional since
I guess I don't really understand it. Does one recognize one's pains when one has them?
Or does one just have them? If I can indeed recognize good instances of MY PAIN, I
suppose it follows that I have the concept PAIN. Does it follow, as compositionality
would require if PAIN is a recognition concept, that I can also recognize good instances
of YOUR PAIN7
Hard cases make bad laws. Sensation concepts are too hard for me.
Chapter 5
There Are No Recognitional Concepts-Not Even
RED, Part 2: The Plot Thickens
Introduction: The Story 'til Now
Some of the nastiest problems in philosophy and cognitive science are
either versions of, or live nearby, what I'll call question Q:
Q: What are the essential (constitutive) properties of a linguistic
expression qua linguistic?
Here are some currently live issues to which I suppose (and to which
I suppose I suppose untendentiously) an answer to Q would provide
the key:
What do you have to learn (know, master) to learn (know, master)
a linguistic expression (concept)? Variant: What are the "possession
conditions" for a linguistic expression (concept)7
1
What is the principle of individuation for linguistic expressions?
What makes two linguistic tokens tokens of the same linguistic
type1
Suppose G is the grammar of language L and E is a lexical expres-
sion in L (roughly, a word or morpheme). What sort of information
about E should G contain?
What's the difference between linguistic and "encyclopaedic"
knowledge?
What belongs to the "mental representation" of a linguistic
expression as opposed to the mental representation of its denotation?
Assume that some of the inferences that involve a lexical item are
constitutive. What distinguishes these constitutive inferences from
the rest7
Which of the inferences that involve a lexical item are analytic?
Assume that some lexical expressions have perceptual criteria of
application (roughly equivalent: Assume that some lexical items
express "recognitional" concepts). Which expressions are these?
Under what conditions is a "way of telling" whether an expression
applies constitutive of the identity of the expression?
SO ChapterS
These are all interesting and important questions, and you will be
unsurprised to hear that I don't know how to answer them. I do, however,
have a constraint to offer which, I'll argue, does a fair amount of work; in
particular, it excludes many of the proposed answers that are currently
popular in philosophy and cognitive science, thereby drastically narrow-
ing the field. Here's the constraint: Nothing is constitutive of the content
of a primitive linguistic expression except what it contributes to the con-
tent of the complex expressions that are its hosts; and nothing is con-
stitutive of the content of a complex expression except what it inherits
from (either its syntax or) the lexical expressions that are its parts.
Short form (Principle P): The constitutive properties of a linguistic
expression qua linguistic include only its compositional properties.
2
Principle P can't, of course, be a sufficient condition for content con-
stitutivity. Suppose, for example, that all cows are gentle. Then, all brown
cows are gentle a fortiori, so "cow" contributes gentle to its (nonmodal)
hosts, and ''brown cow" inherits gentle from its constituents. It doesn't
follow-and, presumably, it isn't true-that gentle is constitutive of the
content of either ''brown," "cow," or ''brown cow." The situation doesn't
change appreciably if you include modal hosts, since not all necessary
truths are analytic. That is, they're not all constitutive of the content of
the expressions that enter into them. "Two" contributes prime to "two
cows"; necessarily, two cows is a prime number of cows. But I suppose
that prime isn't part of the lexical meaning of "two." Not, anyhow, if con-
cept posession requires the mastery of whatever inferences are content-
constitutive.
Still, P is a serious constraint on a theory of content, or so I maintain.
For example: If P is true, then probabilistic generalizations can't be part of
lexical content. Suppose that if something is a cow, then it is probably
gentle. It doesn't follow, of course, that if something is a brown cow, then
it is probably gentle, so "cow" doesn't contribute probably gentle to its
hosts. Mutatis mutandis for other probabilistic generalizations; so proba-
bilistic generalizations aren't part of lexical content.
Likewise, chapter 4 argued that if you grant principle P it follows,
on weak empirical assumptions, that the epistemic properties of lexical
expressions (e.g., the criteria for telling whether they apply) can't be
among their essential properties. The argument went like this: Suppose
that for modifier A, W is the way of telling whether A applies; and that
AN is a complex expression containing the head N with A as its modifier;
and that w is the way of telling whether AN applies. Then, by principle
P, W must be part of W*; the way of telling whether A applies must be
part of the way of telling whether AN applies.
3
Arguably this works
fine for words like, for example, "triangle" with respect to hosts like,
There Are No Recognitional Concepts, Part 2 51
for example, "red triangle," since if counting the sides is a good way of
telling whether something is a triangle, it's likewise part of a good way
of telling whether it's a red triangle. Nothing is either a triangle or a red
triangle unless it's got three sides. But the suggestion that hosts inherit
ways of telling from their constituents, though arguably it's okay for
"red" in "red triangle," doesn't work, even prima facie, in the general case;
for example, it doesn't work for "fish" in "pet fish."
4
Finding out whether
it lives in a stream or lake or ocean is a good way of telling whether
something is a fish; but it's a rotten way of telling whether it's a pet fish.
Pet fish generally live in bowls. It follows, if principle P is true, that being
a way of telling whether an expression applies is not an essential property
of that expression. All forms of semantic empiricism to the contrary
notwithstanding.
All this, I continued to argue in the earlier paper, should strike you
as unsurprising. A way of telling whether an expression E applies to
an object 0 is, at best, a technique (procedure, skill, or whatever) which
allows one to tell whether E applies to 0, given that 0 is a good instance of
E and the circumstances are favorable. This ~ u s t be right if there's to be any
hope that ways of telling are possession conditions. I guess I know what
"triangle" means. But it's certainly not the case that I can tell, for an
arbitrary object in an arbitrary situation, whether "triangle" applies to it.
(Consider triangles outside my light cone.) At best, my knowing what
"triangle" means is (or requires) my knowing how to apply it to a good
instance of triangles in circumstances that are favorable for triangle
recognition. I don't think anybody disputes this. And, I don't think that
anybody should.
But now, the property of being a good instance doesn't itself compose.
What's a good instance of a fish needn't be a good instance of a pet fish,
or vice versa. For that matter, what's a good instance of a triangle needn't
be a good instance of a red triangle, or vice versa. That goodinstancehood
doesn't compose is, I think, the ineliminable fly in the empiricist's
ointment.
Notice, crucially, that goodinstancehood's not composing does not mean
that "pet fish" or "red triangle" aren't themselves compositional. To the
contrary, 0 is a pet fish iff it's a pet and a fish, and 0 is a red triangle iff
it's red and a triangle; "pet fish" and "red triangle" are thus as composi-
tional as anything can get. What it means, rather, is that the epistemic
properties of lexical items aren't essential to their identity qua linguistic.
Not, anyhow, if principle Pis true and only the compositional properties
of an expression are among its constitutive properties.
The constitutive properties of an expression include only the ones it contrib-
utes to its hosts. But, in the general case, expressions don't contribute their good
52 ChapterS
instances to their hosts (being a good instance of a fish isn't necessary for being a
good instance of a pet fish). Since "criteria" (and the like) are ways of recogniz-
ing good instances, it follows that criteria (and the like) aren't constitutive prop-
erties of linguistic expressions. I think that's a pretty damned good argument
that the epistemic properties of lexical items aren't constitutive of their
identity qua linguistic. However, I have shown this argument to several
philosophical friends who disagree. They think, rather, that it's a pretty
damned good argument that principle P can't be true.
5
In particular, so the
reply goes, epistemic properties, such as having some particular criteria
of application, are essential to morphosyntactically primitive linguistic
expressions (like "red" and "Ash") but not to their hosts (like "red triangle"
and "pet Ash") even in cases where the hosts are semantically compositional.
6
If
this reply is right, then a fortiori, the constitutive properties of a linguistic
expression can't be among the ones that its hosts inherit.
I was, for reasons that the chapter 4 elaborated, surprised to hear this
suggestion so widely endorsed. In particular, I argued like this: Consider
any property P that is constitutive of E but not inherited from E by its
hosts; I'll call such a p r o p e r t y ~ "extra." If E has such an extra property,
then, presumably, it would be possible for a speaker to satisfy the pos-
session conditions for a complex expression containing E without satisfy-
ing the possession conditions for E itself: Somebody who has learned the
linguistically essential properties of "pet Ash," for example, need not have
learned the linguistically essential properties of "pet" or "Ash." For, by
assumption, the mastery of "pet" and "Ash" requires an ability to recog-
nize good instances of each in favorable circumstances; whereas, again by
assumption, the mastery of "pet Ash" requires neither of these abilities
(not even if it does require an ability to recognize good instances of pet
Ash in favorable circumstances)?
But, I supposed, it is something like true by de6nition that mastering a
complex expression requires mastering its constituents, since, after all,
constituents are by de6nition parts of their hosts. So, to say that you
could master "pet Ash" without mastering "pet" (or, mutatis mutandis,
"red triangle" without mastering "red") is tantamount to saying that "pet"
isn't really a constituent of "pet Ash" after all; which is, in turn, tantamount
to saying that "pet Ash" is an idiom. Which, however, "pet Ash" patently
is not. So I win.
Now, there is a reply to this reply. (''Dialectics," this sort of thing is
called.) One could just stipulate that recognitional capacities (or, indeed,
any other sort of extra that you're fond of) are to count as constitutive of
the primitive expressions that they attach to even though they are not
inherited by the hosts of which such primitives are constituents. To which
reply there is a reply once again; namely, that explanation is better than
There Are No Recognitional Concepts, Part 2 53
stipulation. Whereas principle P explains why meeting the posession con-
ditions for a complex expression almost always
8
requires meeting the
possession conditions for its constituents, the proposed stipulation just
stipulates that it does; as does any account of constituency that allows
primitive expressions to have extras.
To which there is again a reply: Namely, that you can't expect a theory
to explain everything; and, given a forced choice between empiricism and
the identi6cation of the constitutive properties of an expression with its
compositional properties, one should hold on to the empiricism and give
up principle P. If that requires a revision of the notion of constituency, so
be it. That can be stipulated too, along the lines: A (syntactic) part of a
complex host expression is a constituent of the expression only if it con-
tributes all of its semantic properties to the host except (possibly) its epis-
temic ones.
Thus far has the World Spirit progressed. I don't, myself, think well of
philosophy by stipulation; but if you don't mind it, so be it. In this paper,
I want to float another kind of defense for principle P; roughly, that
the leamability of the lexicon demands it. I take it that these two lines
of argument are mutually compatible; indeed, that they are mutually
reinforcing.
Compositonality and Learnability
My argument will be that, given the usual assumptions, leamability
requires that primitive expressions (lexical items) have no extras. A for-
tiori, it can't be that criteria, recognitional capacities, etc., are constitutive
of primitive' linguistic expressions but not inherited by their hosts.
The "usual assumptions" about leamability are these:
(i) The languages with whose leamability we are concerned are
semantically in6nite (productive); that is, they contain in6nitely
many semantically distinct expressions.
(ii) A theory of the leamability of a language is a theory of the
leamability of the grammar of that language. Equivalently, a learn-
ability theory has the form of a computible function (a '1earning
algorithm") which takes any adequate, 6nite sample of L onto a
correct grammar G of L. I assume, for convenience, that G is unique.
(iii) A learning algorithm for Lis "adequate" only if thegrammar
of L that it delivers it tenable. A grammar G of L is tenable iff L
contains not more than 6nitely many expressions that are counter-
examples to G (and all of these 6nitely many counterexamples are
idioms. See fn.8).
54 Chapter 5
Comments:
Assumption (i) is inessential. As usual, systematicity would do as
well as productivity for any serious polemical purposes. (See Fodor
and Pylyshyn, 1988.)
Assumption (ii) is inessential. It's convenient for the exposition to
assume that learning a language is learning a theory of the language,
and that the relevant theory is a grammar of the l a n ~ g e . But
nothing turns on this. If, for example, you hold that language learn-
ing is '1earning how" rather than '1earning that," that's perfectly
okay for present purposes; the kind of argument I'm going to run
has an obvious reformulation that accomodates this view. I myself
am committed to the idea that learning the semantics of the lexicon
is neither '1eaming that" nor '1earning how"; it's becoming causally
(nomologically) connected, in a certain information-engendering
way, to the things that the item applies to. That view of language
learning is also perfectly okay with the present line of argument, I'm
glad to report.
Though I've borrowed the style of i-iii &om learnability theory,
it's important to see that the intuition they incorporate is really
quite plausible independent of any particular theoretical commit-
ment. Suppose that as a result of his linguistic experience, a child
were to arrive at the following view (subdoxastic or explicit) of
English: the only well-formed English sentence is ''Burbank is damp."
Surely something has gone wrong; but what exactly7 Well, the
theory the child has learned isn't tenable. English actually offers
infinitely many counterexamples to the theory the child has arrived
at. For not only is "Burbank is damp" a sentence of English, but so
too are "Burbank is damp and dull," "The cat is on the mat," ''That's
a rock," and so on indefinitely. The core of i-iii is the idea that a
learning algorithm that permits this sort of situation is ipso facto
inadequate. A technique for learning L has, at a minimum, to deliver
an adequate representation of the productive part of L; and to a cor-
rect representation of the productive part of a language there cannot
be more than finitely many counterexamples among the expressions
that the language contains.
Assumptions (i-iii) constrain language learning rather than concept
learning. I've set things up this way because, while I'm sure that
lexicons are learned, I'm not at all sure that concepts are. What
I'm about to offer is, in effect, a transcendental argument &om the
premise that a lexicon is learnable to the conclusion that none of the
properties of its lexical items are extras. Now, historically, tran-
scendental arguments &om the possibility of language learning are
There Are No Recognitional Concepts, Part 2 55
themselves typically vitiated by empiricist assumptions: In effect,
they take for granted that word learning involves concept acquisi-
tion (that, for example, the questions '1tow does one learn [the word]
'red'?" and '1tow does one learn [the concept] REDr' get much the
same answer). This begs the question against all forms of conceptual
nativism; and I'm inclined to think that some or other sort of con-
ceptual nativism is probably true. (For discussion of the methodo-
logical situation, see Fodor and Lepore, 1992.)
''Transcendental arguments from language learning are typically vitiated
by empiricist assumptions" might do well as the epitaph of analytical
philosophy. I propose, therefore, to consider seriously only those tran-
scendental conditions on language learning that persist even if it's assumed
that concept acquisition is prior to and independent of word learning. That is, I
will accept only such transcendental constraints as continue to hold even
on the assumption that learning the lexicon is just connecting words with
previously available concepts. I take it that constraints of that sort really
must be enforced. Even people like me who think that RED is innate agree
that "red" expresses RED has to be learned.
So, here's the argument at last.
Consider, to begin with, "pet fish"; and suppose that good instances of
fish are R is part of the lexical entry for "fish": According to this lexicon,
being R is part of the content that the word "fish" expresses (it's, if you
prefer, part of the concept FISH). Finally, suppose that a specification of R
is not among the contributions of "fish" to "pet fish" (being R is not part of
the concept PET FISH), so R is an extra within the meaning of the act.
Then, on the one hand, the lexicon says that it's constitutive of "fish" that
being R is part of what tokenings of "fish" convey; but, on the other hand,
the tokenings of "fish" in "pet fish" do not convey this. So "pet fish" is a
counterexample to this lexicon.
Notice that the leamability of "fish"is not impugned by these assump-
tions so far; not even if principle P is true.
9
Tenability requires that English
should offer not more than finitely many exceptions to the lexical representa-
tion of its primitive expressions. But, so far, all we've got is that English
offers one exception to the lexical entry that says that "fish" conveys being
R; namely, "pet fish," which typically does not convey being R. The most
we're entitled to conclude is that, on the present assumptions, the child
couldn't learn "fish" from data about "pet fish." Which maybe is true
(though, also, maybe it's not).
However, an infinity of trouble is on the way. For it's not just the case
that good instances of fish needn't be good instances of pets; if's also the
case that good instances of big pet fish needn't be (for all I know, typi-
cally aren't) good instances of pets or of fish or of pet fish. Likewise, good
56 ChapterS
instances of big pet fish that are owned by people who also own cats
needn't be (for all I know, typically aren't) good examples of fish, or of pet
fish, or of big pet 6sh, or of big pet fish that are owned by people who
also own cats and who live in Chicago.
And so on, forever and forever, what with the linguistic form (A*N)N
10
being productive in English. The sum and substance is this: The good
instances of the nth expression in the sequence A* N are not, in general,
inherited from the good instances of expression n - 1 in that sequence;
and they are not, in general, contributed to the good instances of expres-
sion n + 1 in that sequence. Informally: you can be a good instance of N
but not a good instance of AN; "pet fish" shows this. But, much worse,
you can be a good instance of AN but not a good instance of A(AN)N;
and you can be a good instance of A(AN)N but not a good instance of
AN. Good instances don't, as I'll sometimes say, distribute from modifiers
to heads or from heads to modifiers, or from modifiers to one another.
And their failure to do so is iterative; it generates an infinity of counter-
examples to the compositionality of goodinstancehood.
11
Nothing in this
depends on the specifics of the example, nor is the construction "A*N"
semantically eccentric. Adverbs, relatives, prepositional phrases, and
indeed all the other productive forms of modification, exhibit the same
behavior mutatis mutandis. Thus good instances of triangle needn't be
good instances of red triangle; and good instances of red triangle needn't
be good instances of red triangle I saw last week; and good instances of red
triangles I saw last week in Kansas needn't be good instances of things
I saw last w e e ~ or of triangles, or of red triangles, or of things I saw last
week in Kansas, etc. The hard fad: Goodinstancehood doesn't distribute and
is therefore not productive.
Every primitive expression has an in6nity of hosts to which it fails to
contribute its good instances (mutatis mutandis its sterotype, its proto-
type, etc.). So, a grammar that takes the ability to recognize good instances
to be a possession condition for primitive items faces an infinity of
counterexamples. As far as I know, this point has been missed just about
entirely in the discussion of the "pet fish" problem in the cognitive
science literature (to say nothing of the philosophical literature, where the
problems that compositionality raises for epistemic theories of meaning
have hardly been noticed at all). In consequence there's been a tendency
to think of the semantics of "pet fish," "male nurse," and the like as
idiosyncratic. It's even been suggested, from time to time, that such
expressions are idioms after all. But, in fad, the pet fish phenomenon is
completely general. That typical pet fish aren't typical pets or typical fish
is a fad not different in kind from the fad that expensive pet 6sh are,
practically by de6nition, not typical pet 6sh; or that very, very large red
triangles aren't typical red triangles (or typical red things; or typical
There Are No Recognitional Concepts, Part 2 57
triangles). To repeat the point: it's not just that there are occasional excep-
tions to the compositionality of goodinstancehood. It's that there is an infin-
ity of exceptions-which, of course, assumption (iii) does not permit.
Short form of the argument: The evidence from which a language is
learned is, surely, the tokenings of its expressions. So, if a lexical expres-
sion is to be learnable, its tokenings must reliably manifest its constitutive
properties;
11
that's just a way of saying what tenability demands of learn-
ability. But tokenings of A(AN) do not, in general, reliably manifest the
constitutive properties of AN on the assumption that the epistemic prop-
erties of A (or of N) are constitutive. Nothing about tokenings of "pet
fish" reliably manifests the properties of good instances of fish (or of
pets) tout court; nothing about triangles I saw in Kansas reliably manifests
the properties of good instances of triangles tout court. And so on for
infinitely many cases. So either tenability doesn't hold or the epistemic
properties of lexical expressions aren't constitutive. Q.E.D.
Very shorl form of the argument: Assume that learning that fish typically
live in lakes and streams is part of learning "fish." Since typical pet fish
live in bowls, typical utterances of "pet fish" are counterexamples to the
lexical entry for "fish." This argument iterates (e.g., from "pet fish" to ''big
pet fish" etc.). So a lexicon that makes learning where they typically live
part of learning "fish" isn't tenable. Since the corresponding argument
can be run for any choice of a "way of telling" (indeed, for any "extra"
that you choose), the moral is that an empiricist lexicon is ipso facto not
tenable. Q.E.D. again.
So, maybe we should give up on tenability? I really don't advise it. Can
one, after all, even make sense of the idea that there are infinitely many
sentences of English (infinitely many sentences generated by the correct
grammar of English) which, nevertheless, are counterinstances to the
mental representation of English that its speakers acquire? What, then,
would their being sentences of English consist in? Deep metaphysical
issues (about "Platonism" and such) live around here; I don't propose to
broach them now. Suffice it that if you think the truth makers for claims
about the grammars of natural languages are psychological (as, indeed,
you should), then surely you can't but think that the grammars that good
learning algorithms choose have to be tenable.
But I prefer to avoid a priorism whenever I can, so let's just assume that
being a tenable theory of L and being a well-formed expression of L are not
actually interdefined. On that assumption, I'm prepared to admit that
there is a coherent story about learning English according to which recog-
nitional capacities for fish are constitutive of the mastery of "fish" but not
of the mastery of "pet fish" or of any other expression in the sequence A
fish. That is, there's a coherent story about leanability that doesn't require
tenability, contrary to assumption (iii) above.
58 Chapter 5
Here's the story: The child learns "fish" (and the like) only from atomic
sentences ("that's a fish") or from such complex host expressions as hap-
pen to inherit the criteria for fish. (Presumably these include such expres-
sions as: "good instance of a fish," "perfectly standard example of a fish,"
and the like.) Occurrences of "fish" in any other complex expressions are
ignored. So the right lexicon is untenable (there are infinitely many occur-
rences of "fish" in complex expressions to which "fish" does not contrib-
ute its epistemic properties). But it doesn't follow that the right lexicon is
unlearnable. Rather, the lexical entry for "fish" is responsive to the use of
"fish" in atomic sentences and to its use in those complex expressions
which do inherit its epistemic properties (and only to those).
13
But while I'm prepared to admit that this theory is coherent, I wouldn't
have thought that there was any chance of its being true. I'm pretty sure
that tenability is de facto a condition for a learning algorithm for a pro-
ductive language. Here are my reasons.
I ~ s t said that, in principle, you could imagine some data about complex
A* N expressions constraining the lexical entry for N (and A) even if the
lexicon isn't required to be tenable. The criteria of application for "per-
fectly standard example of a fish" are presumably inherited from the
criteria of application for "fish." But, though this is so, it's no use to the
child since he has no way to pick out the complex expression tokens that
are relevant from the ones that aren't. The criteria for "perfectly standard
fish" constrain the lexical entry for "fish"; the criteria for "pet fish" don't.
But you have no way to know this unless you already know a lot about
English; which, of course, the child can't be supposed to do.
What it really comes down to is that if the lexical representations that
the learning procedure chooses aren't required to be tenable, then it must
be that the procedure treats as irrelevant all data except those that derive
from the behavior of atomic sentences. These are the only ones that can be
un-question-beggingly relied on not to provide counterexamples to the
correct entry. If they typically live in streams or oceans is specified in the
lexical entry for "fish," that has to be learned from tokenings of sentences
which reliably manifest that typical fish live in streams or oceans. And, if
you don't know the semantics, so all you've got to go on is form, "a is a
fish" is the only kind of sentence that can be relied on to do so. So it's the
only kind of sentence whose behavior the child can allow to constrain his
lexical entry for "fish."
You might put it like this: Compositionality works "up" from the
semantic properties of constituents to the semantic properties of their
hosts. But learnability works "down" from the semantic properties of
hosts to those of their constituents. If the learning procedure for L is
required to choose a tenable theory of L, then, with only a finite number
of (idiomatic) exceptions, every host can (potentially) provide data for
There Are No Recognitional Concepts, Part 2 59
each of its lexical constituents. If, however, the learning procedure is not
required to chose tenable grammars, then it has to filter out (ignore) the
data about indefinitely many host constructions. The only way it can do
this reliably is by a constraint on the syntax of the data sentences since, by
assumption, their semantics isn't available; it's what the algorithm is trying
to learn. And since, qua "extra," the property of being a good instance
doesn't distribute from heads to modifers, the only syntactic constraint
that an empiricist learning algorithm can rely on is to ignore everything
but atomic sentences.
So then: If a representation of goodinstancehood is an extra, English
couldn't be learned from a corpus of data that includes no syntactically
atomic sentences. Notice that this is quite different from, and much
stronger than, the claim that the child couldn't learn his language from
a corpus that includes no demonstrative sentences. I would have thought,
for example, that routine employments of the method of differences
might enable the learning procedure to sort out the semantics of "red"
given lots and lots of examples like "this is a red rock," "this is a red
cow," "this is a red crayon." etc., even if a child had no examples of "this
is red" tout court. But not so if mastering "red" involves having a recog-
nitional capacity for red things. For even if, as a matter of fact, red rocks,
red cows, and red crayons are good instances of red things, the child has
no way of knowing that this is so. A fortiori, he cannot assume that typical
utterances of "red rock," "red cow," etc. are germane to determining the
epistemic clauses in the lexical entry for "red."
Here's the bottom line. Chapter 4 argued that an empiricist can't
explain why everybody who understands "AN" also understands "A"; he
can, at best, stipulate that this is so. The present argument is that, since
epistemic properties tum on the notion of a good instance, and since
being a good instance doesn't distribute from modifiers to heads or vice
versa, if you insist on epistemic properties being specified in the lexicon
you will have to treat them as extras. But the learnability of the lexicon is
now in jeopardy; for the learning procedure for a lexical item has, some-
how, to ignore indefinitely many expressions that contain it. In practice
this means ignoring everything except atomic expressions, so the empiri-
cist is committed to an empirical claim which is, in fact, very not plausible:
You can only learn a language from a corpus that contains syntactically
atomic sentences. There is, to repeat, no reason in the world to suppose
that this consequence is true. Do you really want to be bound to a theory
of lexical content which entails it? And, anyhow, what business has a
metaphysical theory about meaning got legislating on this sort of issue?
However, I suppose most of the readers of this stuff will be philos-
ophers, and philosophers like demonstrative arguments. And the argu-
ment I've given isn't demonstrative, since as it stands, it's compatible with
60 Chapter 5
a leamability theory that assumes that there are (adequately many) atomic
sentences in the corpora from which language is learned. Well, I can't give
you a demonstrative argument; but I can give you a diagnosis, one that I
think should satisfy even a philosopher. To wit:
''Being able to tell" whether an expression applies is being able to rec-
ognize good instances. But being able to recognize a good instance of
N doesn't mean being able to recognize a good instance of AN. That's
because your ability to recognize a good instance of N may depend on
good instances of N being R; and, since goodinstancehood doesn't com-
pose, t h e r e ' ~ no guarantee that if good instances of N are R, then good
instances of AN are also R. That is, as I remarked above, the empiricist's
bane. Well, but just why doesn't goodinstancehood compose? Not, I think,
because of some strange or wonderful fact about semantic properties that
lexical items fail to share with their hosts, but rather because of a boring
and obvious fact about typicality. What makes something a typical mem-
ber of the set of Xs needn't be, and generally isn't, what makes something
a typical member of some arbitrary sub- (or super-) set of the Xs. And even
when it is, it's generally a contingent fact that it is; a fortiori, it isn't a
necessary truth that it is; a fortiori, it isn't a linguistic truth that it is, since,
I suppose, linguistic truths are necessary whatever else they are. Whether
being red makes a triangle typical of the kind of triangles they have in
Kansas depends on brute facts about what kinds of triangles they have in
Kansas; if you want to know, you have to go and look. So it's perfectly
possible to know what makes something a typical X, and to know your
language, and nonetheless not have a clue what makes something a typical
member of some sub- (or super-) set. What makes a such and such a good
example of a such and such just isn't a question of linguistics; all those
criteriologists and all those paradigm case arguments to the contrary
notwithstanding.
It looks alright to have recognitional (or other epistemic) features
that are constitutive of terms (or of concepts) until you notice that being
a good instance isn't compositional, whereas the constitutive (a fortiori,
the semantic) properties of linguistic expressions have to be. There is, in
short, an inherent tension between the idea that there are recognitional
terms or concepts and the idea that the semantics of language or thought
is productive. That this sort of problem should arise for empiricists in
semantics is perhaps not very surprising: Their motivations have always
had a lot more to do with refuting skeptics or metaphysicians than
with solving problems about linguistic or conceptual content. Empiricist
semanticists took their empiricism much more seriously than they took
their semantics; it was really justification that they cared about, not lan-
guage or thought. Now all that playing fast and loose with the intentional
has caught up with them. Goodl It's about time I
There Are No Recognitional Concepts, Part 2 61
I'm afraid this has been a long, hard slog. So I'll say what the morals are
in case you decided to skip the arguments and start at the end.
The methodological moral is: compositicmality, learnability, and lexical
entry need to take in one another's wash. Considerations of learnability
powerfully constrain how we answer the galaxy of questions around Q.
And considerations of compositionality powerfully constrain the con-
ditions for learnability.
The substantive moral is: All versions of empiricist semantics are false.
There are no recognitional concepts; not even RED.
Notes
1. For all the cases that will concern us, I'll assume that concepts are what linguistic
expressions express. So I11 move back and forth between talking about concepts and
talking about words as ease of exposition suggests.
2. For our present purposes, Principle P can be taken to be equivalent to the "premise P'' of
chapter 4. The difference, in case anyone cares, is that whereas the current version con-
strains concept constitution, the other constrains concept possession. I suppose that what-
ever is required for having a concept should tum out to be constitutive of the concept,
but perhaps not vice versa. If so, then some, but not all of the constitutive properties of
a concept will correspond to what chapter 4 called the "satisfiers for'' the concept.
3. Of course, there will generally be more than one way of telling whether an expression
applies; but there are obvious ways of modifying the point in the text to accomodate
that.
4. It could be, of course, that this is because "red" and "triangle'' express bona fide recogni-
tional concepts but "pet" and "fish" do not. Notoriously, there isn't a lot of philosophical
consensus about what words (or concepts) are good candidates for being recognitional.
But, in fad, it doesn't matter. As will presently become clear, however the recognitional
primitives are chosen, each one has infinitely many hosts to which it doesn't contribute
its proprietary recognition procedure.
S. Philosophical friends who have e n d ~ r s e d this conclusion in one or another form, and
with more or less enthusiasm, include: Ned Block. Paul Horwich, Brian Loar, Chris Pea-
cocke, and Stephen Schiffer. That's a daunting contingent to be sure, and I am grateful to
them for conversation and criticism. But they're wrong.
6. A variant of this proposal I've heard frquently is that, whereas the criterial and denota-
tional properties of primitive expressions are both constitutive, only the denotational
properties of primitive expressions are compositional; i.e., these are the only semantic
properties that primitive expressions contribute to their hosts.
It's more or less common ground that denotation (instancehood, as opposed to good-
instancehood) does compose. The (actual and possible) red triangles are all found in the
intersection of the (actual and possible)" things "red" denotes and the (actual and possi-
ble) things "triangle" denotes; the pet fish are all found in the intersection of the things
"pet" denotes with the things that "fish" denotes. And so on. It's thus in the cards that if,
as I argue, compositionality rules out epistemic properties as linguistically constitutive,
the default theory is that the constitutive semantic properties of words (and concepts)
are their satisfaction conditions.
7. Notice that the argument now unfolding is neutral on the parenthetical caveat, i.e., it
is not committed either way with respect to what Schiffer calls the "Agglomerative
Principle." See chapter 4.
62 Chapter 5
8. That is, always excepting idioms. This caveat isn't question-begging, by the way, since
there are lots of independent tests for whether an expression is an idiom (e.g., resis-
tance to syntactic transformation, to say nothing of not entailing and being entailed by
its constituents).
9. On the other hand, if P is true, then this lexicon implies that "pet fish" is an idiom.
10. "(A*N)N" is a schema for the infinite sequence of (phrasal) nouns: (AtN)N, (A2N)N, etc.
11. Notice that the point I'm making does not depend on modal or "counterfactive" adjec-
tives ("probably,""fake," etc.), where it's well known that AN can be compatible with, or
entail, not-N.
Notice, likewise, that though for convenience I talk of goodinstancehood being not
compositional, exactly parallel considerations show that constraints on good instance-
hood aren't. The requirements for being a good instance of N aren't, in general, inherited
by the requirements for being a good instance of AN; and so forth. This is a special case
of the intertranslatability of stereotype-like theories of concepts (or lexical meanings)
with exemplar-like theories.
12. It isn't good enough, of course, that the data from which an expression is learned should
merely be compatible with its having the constitutive properties that it does. That "pet
6sh" is said of things that typically live in bowls is compatible with "fish" being said of
things that typically don't; for it's contingent, and neither a priori nor linguistic, that
people don't generally keep good instances of fish as pets.
13. Notice that the present proposal is not that the child learns that only primitive expres-
sions exhibit the constitutive epistemic properties. ''Typical triangle" and the like show
that that's not true. The relevant constraint is (not on what the child learns but) on what
. data he learns it from.
Chapter 6
Do We Think in Mentalese7 Remarks on Some
Arguments of Peter Carruthers
For better or worse, I am much inclined to the view that the vehicle of
more or less al}l thought is an innate, nonnatural language. It has become
the custom to call this putative language in which one thinks Mentalese,
and I shall do so in what follows. I'm also much inclined to the view that
the semantics of Mentalese is informational and atomistic-infonnational
in the sense that Mentalese symbols mean what they do in virtue of their
causal/nomic connections to things in the world; and atomistic in the sense
that the metaphysically necessary conditions for any one Mentalese sym-
bol to mean what it does are independent of the metaphysically necessary
conditions for any other Mentalese symbol to mean what it does.
Actually, Carruthers (1996; all page references in this chapter are to this
book) agrees with much of this. In particular, he thinks that the vehicle of
-thought is linguistic, and that a lot of concepts are probably innate. But he
disagrees on two essential points: He holds that all conscious thinking is
done in (as it might be) English;
2
so English is the language of conscious
thought. And he denies that English has an atomistic semantics. Ergo, he
denies that the language of conscious thought does.
Refreshingly, Carruthers doesn't claim that he has knock-down argu-
ments for any of this; he holds only that his is the view of preference
ceteris paribus. In particular, he takes it to be introspectively plausible that
one thinks in English; and, if there are no decisive reasons for supposing
that one doesn't, we ought to accept the introspectively plausible view.
"All I need to show is that Fodor has no definitely convincing argument
for the thesis that the language of all thinking is Mentalese, in order for
the introspective thesis to win by default" (134). Accordingly, much of
the first half of Carruthers's book is given to arguments that are supposed
to show that the arguments that are supposed to show that we think in
Mentalese aren't decisive.
3
These are the arguments I propose to discuss
in what follows.
Not, however, that I disagree with Carruthers's conclusion. It's true that
the arguments that we think in Mentalese aren't decisive. This is, quite
generally, not the kind of area in which decisive arguments are available;
64 Chapter 6
rather, it's the kind of area in which one has had a good day if one has
managed not to drown. But I do think the arguments, for the view that
one thinks in Mentalese are pretty plausible; indeed, I think they are the
most plausible of the currently available options. And I don't think that
Carruthers is right about what he says is wrong with them.
Carruthers has three major points to make; one is about the relation
between the Mentalese story and the Griceian program in semantics; one
is about the relation between the Mentalese story and the notion of func-
tional definition; and one is about the relation between the Mentalese
story and semantic holism. I propose to consider each of these. First,
however, a brief word or two about the introspection argument itself.
To begin with, it's not so introspectively obvious that it's introspec-
tively obvious that one thinks in English. The psychologist Ken Forster
once told me that when he eavesdrops on his thinking, what he hears is
mostly himself saying encouraging things to himself. "Go on, Ken; you
can do it. Pay attention. Try harder." My own thought processes, for what
it's worth, are no more transparent, though the internal commentary that
goes with them is generally less sanguine: "I can't solve this; it's too hard.
I'm not smart enough to solve this. If Kant couldn't solve this, how can
they possibly expect me to7" and so forth. It could be, of course, that
people's introspections disagree.
Anyhow, as Carruthers is fully aware, it's surely not introspectively
available that what one does when one talks to oneself is thinking;
specifically, it's not introspectively available that what one finds oneself
saying to oneself is what plays the causal/functional role of thought in the
fixation of belief, the determination of action, and the like. Introspection
itself can't show this, for essentially Humean reasons: Post hoc is one
thing, propter hoc is another; causal connectedness can't be perceived, it
has to be argued for. That is as true of the inner sense as it is of the outer.
Finally-though this is a long story, and not one I want to dwell on
here-I sort of doubt that the theory that we think in English can actually
be squared with the introspective data, even if we're prepared to take the
latter at face value. The problem is this: "thinking in English" can't just be
thinking in (sequences of) English words since, notoriously, thought
needs to be ambiguity-free in ways that mere word sequences are not.
There are, for example, two thoughts that the expression "everybody
loves somebody" could be used to think, and, so to speak, thought is re-
quired to choose between them; it's not allowed to be indifferent between
the possible arrangements of the quantifier scopes. That's because, sans
disambiguation, "everybody loves somebody" doesn't succeed in specify-
ing something that is susceptible to semantic evaluation; and susceptibility
to semantic evaluation is a property that thought has essentially. You can,
Do We Think in Mentalese? 65
to be sure, say in your head "everybody loves somebody" while remaining
completely noncommittal as to which quantifier has which scope. That
just shows that saying things in your head is one thing, and thinking
things is' quite another.
I take the moral to be that "everybody loves somebody" isn't, after all,
a possible vehicle of thought. The closest you could come to thinking
in English would be to think in some ambiguity-free regimentation of
English (perhaps in formulas of what Chomsky calls "LF" (roughly, Logi-
cal Form). Maybe, for example, what's in your head when you think that
everybody loves somebody on the interpretation where "everybody" has
long scope, is "every.r some
11
(x loves y)." That (give or take a bit) is
the right sort of linguistic structure to be the vehicle of a thought. But
(dilemma) it's surely not the sort of linguistic structure that is given to
anybody' s introspection; if it were, we wouldn't have needed Frege
to teach us about bound variables. Maybe intuition does say that you
think in word sequences; but there's good reason to think that it's wrong
to say so.
It is, to put the same point a little differently, a defining property of
mental representation theories to hold that the causal role of thoughts in
mental processes is formally determined by the structure of their vehicles.
This is where the representational theory of mind intersects the computa-
tional theory of mind, and I take it to be common ground between me
and Carruthers. There are, however, two different causal roles that
converge on the sequence of words "everybody loves somebody." That
being so, this sequence of words corresponds to two different vehicles
of thought even assuming that we think in English. What distinguishes
the vehicles-what the difference in their causal/computational powers
depends on-is their synt.ax.
4
And the syntax of English isn't intro-
spectively available. Carruthers can have it that we think in English, or
he can have it that we have introspective access to the vehicles of our
thoughts; but he can't have it both ways.
Ambiguity isn't all that's wrong with English sentences as vehicles of
thought, by the way; for sometimes they tell you not too little but too
much. Was what you thought that it was surprising that John arrived early,
or was it that John's arriving early was suprising? If you find that your intro-
spections are indifferent, then either they are unreliable or you mustn't
have been thinking in English since, of course, English distinguishes
between these expressions. Carruthers holds that it's precisely the con-
scious thinking that you do in a natural language. So, if your thought that
it was surprising that John arrived early was conscious, there must be an
(introspectively available) fact of the matter about which sentence form
you thought it in. Well, but is there?
66 Chapter 6
However, let's put the introspection stuff to one side. I'm anxious to
get to Carruthers's arguments against the arguments that we think in
Mentalese.
First argument: Mentalese and the Grice program Suppose that thoughts
are relations to Mentalese sentences, and that the content of a thought
is the meaning of the Mentalese sentence to which it is related. And
suppose not only that Mentalese is a nonnatural language, but that it's
an ontologicaly independent nonnatural language; that is, a Mentalese sen-
tence's meaning what it does doesn't depend on any natural language
sentence meaning what it does. What, on such a view, ought one to
say about the meanings of natural language sentences themselves?
According to Carruthers, the thing for a mend of Mentalese to do
would be to adopt "a Griceian approach to natural-language semantics,
according to which sentence-meaning is to be explained in terms of the
communication of beliefs .... [This] takes the notions of belief and inten-
tion for granted and uses them in explaining the idea of natural language
meaning" (7 4).
However, as Carruthers rightly remarks, "Since Grice wrote, there has
been a rich literature of counterexamples to [the Griceian] account" (77);
and, probably, nobody now thinks that the reduction of the meaning of
English sentences to facts about the communicative intentions of English
speakers-or, for that matter, to any facts about mental states-is likely
to go through. This is just a special case of the general truth that nothing
ever reduces to anything, however hard philosophers may try. So, then,
here's the way Carruthers sees the geography: The Mentalese story about
thought is hostage to the Grice story about natural-language meaning;
and the Grice story about natural-language meaning is a check that prob-
ably can't be cashed; so the Mentalese story about thought is, to that
extent, in trouble.
However: Carruthers is wrong to think that it matters to Mentalese
whether Grice can cash his checks. More generally, it doesn't matter to
Mentalese whether any story about natural-language meaning can be
made to work. More generally still, it doesn't matter to Mentalese
whether there is such a thing as the meaning of a natural-language
expression or (in case this is different) whether natural-language expres-
sions have meanings. What Mentalese cares about-and all that it cares
about that connects it with the Grice program-_ is that natural language
is used to express thought.
What's required for natural language to express thought is that know-
ing a natural language can be identi6ed with knowing, about each of its
sentences, which thought it is used to express. It might tum out to be pos-
sible to parley this condition into a theory of natural-language meaning.
Do We Think in Mentalese7 67
On the face of it, it's plausible enough that whatever expresses thought
must ipso facto have a meaning that is determined by the content of the
thought that it expresses. Cashing that intuition is, in effect, what Grice's
theory of meaning was up to. But-to reiterate-neither the claim that
we think in Mentalese, nor the claim that knowing a natural language is
knowing how to pair its expressions with Mentalese expressions, requires
that this be so.
Suppose, for example, that there is no answer at all to questions like
'What does 'the cat is on the mat' meanr' just as there is no answer at all
to questions like 'What do tables and chairs meanr' or 'What does the
number 3 mean?" English sentences, numbers, and tables and chairs, are,
by this supposition, not the sorts of things in which meaning inheres.
Perhaps only thoughts or mental representations are. It wouldn't follow
that there is no answer to questions like 'What thought is 'the cat is on
the mat' used to expressr' In fact, it's used to express the thought that
the cat is on the mat, whether or not it thereby means that the cat is
on the mat; indeed, whether or not it means anything. To repeat: Neither
thoughts having content, nor the use of language to express thoughts,
depends on language having content. Not, at least, by any argument that
I can think of; certainly not by any of the arguments that Carruthers has
on offer.
But, if that's right, then the Mentalese story about thinking doesn't
depend on the Griceian reduction going through after all. The most that
could follow from the failure of Grice's reduction is that he was wrong to
assume that whatever is used to express thought must mean something. It
would, no doubt, be surprising and rather sad if Grice was wrong about
this, since the intuition that sentences, words, and the like do mean some-
thing would have to be given up,
5
and it is a plausible intuition, as pre-
viously remarked.
6
But, once again, the status of Mentalese wouldn't be in
the least impugned.
Here's a slightly different way to put this point. Grice had a theory of
linguistic communication and a theory of linguistic meaning, and his idea was
somehow to ground the latter theory in the former. Now, Grice's theory
of linguistic communication is practically untendentious, at least for any-
body who is a realist about intentional states. The theory was that you
utter sentences intending, thereby, to indicate or express the content of
the mental state that you're in; and your hearer understands you if he
determines from your utterance the content of the mental state that you
intended it to express or to indicate. I take this theory of communication
to be practically platitudinously true. So far as I know, none of the myriad
counterexamples to the Grice program that you find in the literature is an
objection to his communication theory.
68 Chapter 6
Notice, however, that this communication theory could be true even if
mental states are the only things in the world in which content inheres. In
particular, no commitments at all about natural-language meaning enter
into it. This is not an accident; indeed, it's essential from Grice's point of
view that his account of communication should not presuppose a notion
of natural-language meaning. For his goal was to derive the theory of
meaning from the theory of communication, and he had it in mind that
the derivation should not be circular.
I suppose Mentalese does need Grice's communication theory; or, any-
how, something like it. An acceptable account of thought ought to say
something about how thoughts are expressed; and if thoughts are what
have content in the first instance, then it is natural to suppose that what
communication communicates is the content of thoughts. But though
it needs his theory of communication, Mentalese doesn't need Grice's
theory of natural-language meaning; or, indeed, any theory of natural lan-
guage whatsoever. For the Mentalese story is not just that the content of
thought is prior to natural-language content in order of explanation; the
Mentalese story is that the content of thought is ontologically prior to
natural-language meaning. That is, you can tell the whole truth about
what the content of a thought is without saying anything whatever about
natural-language meaning, including whether there is any. Here, according
to the Mentalese story, is the whole truth about the content of a thought:
it's the content of the Mentalese sentence whose tokening constitutes the
having of the thought. (This, by the way, is about as un-Wittgensteinian
a view of the conceptual geography as it is possible to hold. If the
Mentalese story about the content of thought is true, then there couldn't be
a private language argument. Good. That explains why there isn't one.)
It may be just as well that Mentalese leaves all the questions about
natural language semantics open, because-so it seems to me-it really
might tum out that there is no such thing. Maybe all there is to what
1/cat" means is that it's the word that English speakers use to say what
they are thinking about when they are thinking about cats. That, of
course, isn't semantics in either of the familiar construals of the term: It
doesn't specify either a relation between the word and its meaning or a
relation between the word and what it's true of. It can't, however, be held
simply as a point of dogma that English has a semantics; if you think that
it does, then you will have to argue that it does. Grice failed to say what
natural-language meaning is; and so too, come to think of it, has every-
body else. Maybe natural-language is the wrong tree for a theory of
meaning to bark up.
Bottom line: Carruthers thinks Mentalese is hostage to the Grice pro-
gram. But it's not. So the failure of the Grice program isn't an argument
against the claim that we think in Mentalese.
Do We Think in Mentalese7 69
I tum, briefly, to Carruthers's last two arguments, both of which con-
cern the doctrine that the semantics of Mentalese is atomistic.
To begin with, not every friend of Mentalese accepts this doctrine.
Indeed, the standard view is that the meaning of Mentalese expressions
is, at least partly, determined by their functional (inferential) roles. Car-
ruthers himself holds that this is so, both for Mentalese and for English.
But I have resisted assimilating the Mentalese story about thought to a
functional role theory of meaning. That's because I believe that functional
role semantics is intrinsically holistic and that a holistic semantics is
incompatible with any serious intentional realism.
The argument for the second step is familiar: If is holistic,
then the content of each of your thoughts depends on the content of each
of your others; and, since no two people (indeed no two time slices of the
same person) have all their thoughts in common, semantic holism implies
that there are no shared intentional states. Hence that there are no sound
intentional generalizations. Hence that there are no intentional explan-
ations. But intentional Realism just is the idea that our intentional mental
states causally explain our behavior; so holistic semantics is essentially
irrealistic about intentional mental states. So the story goes.
Carruthers has two points to make against this, neither of which seems
to me to work, but both of which keep cropping up in the recent liter-
ature. Let's have a look at them.
First, though Carruthers admits that there are lots of differences
between people's actual beliefs, he thinks that inferential role semantics,
rightly formulated, can avoid concluding that there are no, or practically
no, belief contents that people share. The idea is that "[the contents of]
propositional attitudes are individuated in terms of their potential [sic]
causal interactions with one another. These . . . may remain the same
across subjects whose actual propositional attitudes may differ .... mhe
same conditionals can be true of them" (111). So, for example, you know
that today is Tuesday, so you infer that tomorrow is Wednesday. I don't
know what day it is today, so I'm agnostic about what day it will be
tomorrow. But that's okay because we do have this hypothetical in com-
mon: if we think today is Tuesday, then we think tomorrow is Wednes-
day. Since, according to the present proposal, the functional role of an
attitude, and hence its content, is to be defined relative to these shared
conditionals, the eccentricity of actual beliefs needn't imply that belief
contents can't be shared. As I say, this sort of idea is getting around
(cf. 1993).
Well, it may help a little, but it surely doesn't help much. For one thing,
belief hypotheticals shift along with the categoricals when indexicals are
involved. It's because I believe that today is Tuesday that I believe that if
I ask you, that's what day you'll say it is. If content identity is relativized
70 Chapter 6
to such hypotheticals, it's going to exhibit just the sensitivity to individ-
ual differences that Carruthers is trying to avoid.
And even where "eternal" propositions are at issue, it's a classic prob-
lem for functional role semantics that, on the one hand, what functional
relations hold between your beliefs depends on what you think is evi-
dence for what; and, on the other hand, what you think is evidence for
what depends, pretty globally, on what else you happen to believe-that
is, on what you categorically believe. You think Clinton is a crook because
you've read that he is in the Times. I haven't read the Times, so we differ
in categorical beliefs (viz., about what it says in the paper), so we can't
share relevant Clinton-belief contents. Thus the old holism. Notice how
little the move to hypotheticals does to help with this. What, for example,
is the status, in the circumstances imagined, of the conditional: "If one has
read the Times, then one thinks that Clinton is a crook." Answer: though
it's true of you by assumption, it's quite likely to be false of me: If I'd read
that about Clinton, dogmatic Democratic that I am, I would have con-
cluded not that he's a crook but that the Times is unreliable. One man's
modus ponens is another man's reductio, as epistemologists are forever
pointing out.
Examples like "if you think it's Tuesday today, then you think it's
Wednesday tomorrow," where belief hypotheticals are plausibly content-
constitutive (and hence insensitive to, inter alia, the contents of con-
tingent categoricals) are exceptional and therefore misleading. Rather,
people who differ in their categorical beliefs are often likely to satisfy
different belief hypotheticals for that very reason.
7
Bottom line: There is a
short ~ o u t e from functionalism about meaning to meaning holism. As far
as anybody knows, there's no way to get around that.
What's wrong with functional role semantics (FRS) is that it wants to
analyze the content of a belief in terms of its inferential (causal) relations;
whereas, plausibly, the direction of metaphysical dependence actually
goes the other way 'round. The content of a belief determines its causal
role; at least, it does in a mind that's performing properly. Holism is
Nature's way of telling FRS that it has the direction of analysis backwards.
None of this is discernably altered if FRS goes hypothetical.
Alright, alright, but doesn't that prove too much? "[Holism] is a problem
for functional individuation generally, not a problem for functional role
semantics in particular. So, unless we are prepared to accept that there
really are no functionally individuated entities, we have as yet been given
no reason for rejecting functional individuation of thoughts" (114). (See
also Devitt, 1996, where this sort of line is pushed very hard.)
Well, but does avoiding the holistic embarrassments of FRS really
require rejecting functional analysis per se? That, to be sure, would be a
Do We Think in Mentalese7 71
steep price; more than even people like me, who hate FRS root and
branch, are likely to be willing to pay.
Notice, to begin with, that the claim that the individuation of such and
suches is functional is, in and of itself, a piece of metaphysics pure and
simple; in and of itself, it need have no semantical implications at all. Sup-
pose, for example, I claim that hearts are functionally individuated; in par-
ticular, that hearts just are whatever it is that pumps the blood. Making
this claim presumably commits me to a variety of metaphysical neces-
sities; if it is true, then ''hearts are pumps" and all of its entailments are
among them. But it commits me to nothing whatever that's semantic. For
example, it leaves wide open all questions about what, if anything, the
word ''heart" means; or about what the principles of individuation for the
concept HEART are; or about what the possession conditions are for
having that concept.
The reason that claiming that hearts are pumps leaves all this open is
that you can't assume-not at least without argument-that if it's meta-
physically necessary that Fs are Gs, then it follows that "F" means some-
thing about being G; or that knowing that Fs are Gs is a possession
condition for the concept F (or for the concept G). Thus, for example: It's
metaphysically necessary that two is the only even prime. It doesn't
follow that "two" means the only even prime or that having the concept
TWO requires having the concept PRIME, etc. Likewise, it's metaphysi-
cally necessary that water is H
2
0. But it doesn't follow that "water''
means H2 0; or that you can't have the concept WATER unless you have
the concept HYDROGEN, etc.
Functionalism about hearts is, to repeat, a metaphysical thesis with,
arguably, no semantical implications whatever. Whereas, the holistic embar-
rassments for FRS arise precisely from its semantical commitments; in particu-
lar, from the constraints that it places on concept possession. The intention of
FRS is that if I is an inference that is constitutive of concept C, then hav-
ing concept C requires in some sense acknowledging I (e.g., it requires
finding I "primitively compelling," or it requires being disposed to draw I
ceteris paribus, or whatever). Holism follows from this construal of con-
cept possession together with the assumption (independently warranted,
in my view) that there is no principled way to distinguish concept-
constitutive inferences from the rest.
But, notice, this is semantic functionalism; it's functionlism about having
the concept HEART (or about knowing what the word ''heart" means).
Semantic functionalism does lead to semantic holism, and by familiar
routes. But ontological functionalism is fully compatible with semantic
atomism. That's because, although functionalist theories of hearts bring all
sorts of metaphysical necessities in their train, they place no constraints
72 Chapter 6
whatever on concept possession. It's just a fallacy to infer the functional indi-
viduation of concepts from the functional individuation of the properties that
they express.
So it's alright to be a functionalist about hearts; at least it's alright for
all the arguments that we've seen so far. But it's not alright to be a func-
tionalist about conceptual content; that leads to holism, hence to unsat-
isfiable constraints on concept possession.
One last point on all of this: It is, for all I know, okay to be a function-
alist about conceptual content if you can, at the same time, contrive to
avoid being a semantic holist. And, of course, an FRS functionalist can
avoid being a semantic holist if he is prepared to accept an analytic/
synthetic distinction; which, in fad, Carruthers does. I don't want to argue
about whether it's okay to philosophize on the assumption of an analytic/
synthetic distinction; everyone has to sleep with his own conscience.
Suffice it that you don't need an analytic/synthetic distinction-or any
semantical views whatever-to be a functionalist about hearts and such;
you just need to be an essentialist. Being an essentialist is quite a different
thing from endorsing an analytic/synthetic distinction (unless there is rea-
son to believe that truths about essences are ipso facto analytic-which
there isn't because they aren't). That's another way to see why functional-
ism about concepts doesn't follow from functionalism about things; and
why the former is so much the more tendentious doctrine. So much, then,
for Carruthers's third argument.
I want to end by returning to a point I've already made in passing.
I don't think that there are decisive arguments for the theory that all
thought is in Mentalese. In fad, I don't think it's even true, in any detail
that all thought is in Mentalese. I wouldn't be in the least surprised, for
example, if it turned out that some arithmetic thinking is carried out by
executing previously memorized algorithms that are defined over public
language symbols for numbers (''Now carry the '2,'" and so forth). It's quite
likely that Mentalese co-opts bits of natural language in all sorts of ways;
quite likely the story about how it does so will be very complicated
indeed by the time that the psychologists get finished telling it.
But here's a bet that I'm prepared to make: For all our philosophical
purposes (e.g., for purposes of understanding what thought content is,
and what concept possession is, and so forth), nothing essential is lost
if you assume that all thought is in Mentalese. Hilary Putnam once
remarked that if you reject the analytic/synthetic distinction, you'll be
right about everything except whether there is an analytic/synthetic dis-
tinction. Likewise, I imagine, in the present case: If you suppose that all
thought is in Mentalese, you'll be right about everything except whether
all thought is in Mentalese. More than that is maybe more than it's rea-
sonable to ask for.
Do We Think in Mentalese7 73
Appendix: Higher-Order Thoughts
I try never to think about consciousness. Or even to write about it. But I
do think there's something peculiar about Carruthers's treatment. I quote
(17 4): "Any mental state M, of mine, is conscious = M is disposed to
cause an activated [i.e., not merely dispositional] conscious belief that
I have M." Now, as Carruthers rightly notes, this formulation would
appear to be circular on the face of it. So he immediately adds "so as to
eliminate the occurrence of the word 'conscious' from the right hand side
of the definition" the following:
Any mental state M, of mine, is conscious = M (level I) is disposed to
cause an activated belief that I have M (level 2), which in tum is disposed
to cause the belief that I have such a belief (level 3) and so on; and every
state in this series, of level n, is disposed to cause a higher-order belief of
level n + 1.
But that can't really be what Carruthers wants. It eliminates the occur-
rence of the word "conscious" from the right hand side alright; but it does
so precisely by leaving it open that each of the higher-order beliefs
involved might be unconscious. And it is explictly Carruthers's view that
having the unconscious n-level belief that one is having a certain n - 1
level thought is not sufficient to guarantee that the n - 1 level thought is
conscious.
As far as I can see, two options are open to neither of which
is satisfactory. He could simply reintroduce "conscious" to qualify each of
the higher level beliefs (e.g., M [level 1] is disposed to cause an activated
conscious belief that I have M [level 2], etc.)-thereby, however, failing to
resolve the problem about circularity. Or he could deny that "conscious"
really occurs in "conscious-level-n belief" for any level higher than 1;
in effect, he could spell "conscious-level-n belief" with hyphens-there-
by, however, raising the hard question why a merely conscious-with-a-
hyphen level n + 1 belief should suffice to make a level-n belief conscious
without a hyphen. Notice, in partirular, that conscious-with-a-hyphen
beliefs can't be assumed to be conscious sans phrase; if they were, the
analysis would lapse, once again, into circularity.
Notes
1. This is to ignore a lot of subtleties; for example, whether we sometimes think in images,
and the like. Such questions may matter in the long run, but not for the project at hand.
2. In what follows ''English" is generally short for ''English or some other natural language"
and "sentence" is generally short for "sentence or some other expression." Nothing hangs
on this except brevity of exposition.
3. Most of the second part of Carruthers's book is about consciousness. I say a little about
that in the Appendix.
74 Chapter6
4. I gather from e-mails that Carruthers thinks what makes the difference between the two
thoughts that everybody loves somebody is not the syntax of the representation that you
token, but something like the fDt1Y vou intend the token. I'm not sure that I understand that.
I don't think that I intend my thoughts at all; I think that I just have them But, in
any event, the suggested doctrine is pretty clearly not compatible with a computational
theory of mind. so I don't want any.
S. I'm assuming that if "the cat is on the mat'' is used to express the thought that the cat is
on the mat, then if it means anything it means that the cat is on the mat.
6. Disquotation would still give a sense to the thought that sentences, words, and the like
mean something; to that extent they would be better off than nwnbers and tables and
chairs, for which conventions of (dis)quotation have not been defined. Cold comfort, if
youaskme.
7. There are, of course, all sorts of other ways in which belief hypotheticals shift with belief
categoricals. I am cynical; I think all politicians are crooks. So (given, as usual, reasonable
rationality) it's true of me that if I think Francis is a politician, then I think Frands is a
crook. You are naive; you think no politicians are crooks. So, it's not true of you that if
you think Francis is a politician then you think Francis is a crook. If the contents of our
beliefs are relativized to these hypotheticals, we get the usual dreary result: Our beliefs
about politicians aren't contradidories.
Chapter 7
Review of A. W. Moore's Points of View
Proust's Swann is obsessed by what he doesn't know about Odette. His
anguish has no remedy; finding out more only adds to what he does
know about her, which isn't what's worrying him. Since Kant, lots of phi-
losophers have suffered from a generalized and aggravated form of the
same complaint. They want to know what the world is like when they
aren't thinking about it; what things are like, not from one or other point
of view, but "in themselves." Or, anyhow, they think that maybe that's
what science aims to know, and they wonder whether it's a project
that makes any sense. They are thus concerned with "the possibility of
objectivity."
Moore's book is a sort of defense of the possibility of objectivity (what
he often calls the possibility of "absolute representations"). He doesn't
argue that objectivity is ever attained or ever likely to be; perhaps not
even in our most scrupulous investigations. All he wants is that the goal
should be coherent. Given the modesty of the enterprise, it's really
shocking the conclusions that he's driven to. 'We are shown that abso-
lute representations are impossible .... [But] ... we do well to remind our-
selves ... that what we are shown is nonsense. Properly to replace 'x
in the schema "A is shown that x' is a quasi-artistic exercise in which one
creates something out of the resonances of (mere) verbiage. There is no
reason whatever why this should not sometimes involve making play
with inconsistency ... " (272)
It's, anyhow, a fascinating story how Moore gets to this. I think, in fact,
that it's a cautionary tale. Moore takes on board, from the beginning and
essentially without argument, the currently received view of meaning in
philosophy. The rest follows as the night the day. What's splendid about
the book is that, unlike most philosophers who share his premises, Moore
is prepared to face the consequents. What's appalling is the tenacity with
which Moore clings to an account of representation from which it follows
that his own philosophical views, inter alia, are nonsense. ('1nter alia"
because, of course, your views and mine are nonsense too.) Not just like
nonsense, mind you, but the real thing: on all fours with "Phlump jing
76 Chapter 7
ax," since "there can be no other reason for an utterance's failing to be a
representation than that certain [of the] words lack meaning" (201). I pro-
pose to trace, briefly, the course of these events. At the end, I'll moralize a
little about how things stand if Moore's book is read, as I think it should
be, as a reductio of his theory of meaning.
The book starts oddly: "[A] representation need not be objective .... [A]
representation can be from a point of involvement. . . . I shall call any
[such] representation ... 'subjective'" (8). This makes representations that
contain indexicals (like "It is humid today") perspectival and subjective,
since "[l]f I wish to endorse an assertion I made yesterday of [this] sen-
tence, I have no alternative but to produce a representation of some other
type ('It was humid yesterday')" (12). This, to repeat, is an odd way to
start. For although indexical sentences are uttered from a point of view (in
the sense that "it's raining" can be true here and false there at one and the
same time; or true then and false now in one and the same place), there
doesn't seem to be anything particularly subjective about them. If what
you want is subjectivity, try "red is warmer than green" or "Callas was
better than Tebaldi."
Indexicality is about content, whereas objectivity is about truth. If you
take indexicality as a paradigm of the perspectival, then it's a mistake to
run "perspectival" together with "subjective." What's interesting about a
sentence like "it's raining" is that the content of the assertion that I make
when I say it depends on where and when I am. What's interesting about
"red is warmer than green" is something quite different; namely, that
(maybe) it can be true for me but false for you, or (maybe) there's no
matter of fact at all about whether it's true. There's, anyhow, certainly a
matter of fact about whether it's raining; so what on earth does whether a
representation is indexical have to do with whether it's objective? So one
wonders as one reads along.
I guess what Moore has in mind is something like this: The content of
an indexical representation depends on its context. But so likewise do the
contents of representations of every other kind since having a representa-
tion means having a language, and ''having a language involves having an
outlook; or, more specifically . . . having a language involves having its
own distinctive outlook" (94). Moore swallows whole not just Wittgen-
stein, but also Whorf and Saussure; which is, perhaps, a little quaint of
him. So the content of a representation, indexical or otherwise, is fixed
partly by its conceptual role, by the inferences it licenses. That one can
only represent things from the point of view of one's inferential commit-
ments is therefore part and parcel of what content is. But representations
from a point of view are ipso facto perspectival, hence ipso facto not
objective. It seems, in short, that semantics puts pressure on epistemology
Review of Moore's Points of View 77
across the board; there is a prima facie conflict between the possibility of
objectivity and the idea that content is holistic. How to resolve this con-
Aid is what Moore's book is really about, though I doubt he'd want to put
the case this way.
You don't have to resolve it, of course; you could try to just live with
it. Quite a lot of philosophers, including some of our most eminent con-
temporaries, seem to think that's indeed the best that you can do. So
Thomas Kuhn famously maintained that, because concepts are implicitly
defined by the theories that endorse them, radically different theories are
ipso facto "incommensurable." There is, to that extent, no such thing as a
rational choice between conflicting theoretical paradigms, and there is no
fact of the matter which, if either, of conflicting paradigms is objectively
correct. Changing your paradigm isn't changing your mind, it's changing
your lifestyle.
Variants of this disagreeable idea are on offer from philosophers with
as little else in common as, say, Wittgenstein, Quine, Goodman, Putnam,
Derrida, and Davidson, among many others. Part of what's wrong with it
is that it's so unstable. Just how major does the disagreement in back-
ground commitments have to be before incommensurability sets in7 And
if, as seems likely, there is no principled answer to this question, isn't it
the moral that all judgment turns out to be more or less subjective? We're
not far from the view that argument and persuasion are mostly power
politics in disguise; class politics on some accounts, sexual politics on
others. Whatever became of the disinterested pursuit of objective truth?
There is, however, a traditional, more or less Kantian way to split the
difference between what's perspective and what isn't; it's the doctrine
called "transcendental idealism," and much of what's most interesting
in Moore's book is about it. It's one thing, Moore says, to suppose that
representation is ipso facto "from an outlook"; it's another to conclude
that representation is ipso facto not objective. The second follows from
the first only on the assumption that apparently conflicting perspectival
representations aren't capable of being integrated from a point of view
that embraces both.
Suppose one takes for granted that representation is, as Moore puts it,
always "of what's there anyway." That is, all of our points of view are
perspectives on the same world, the world to which our beliefs conform
insofar as they are true. True beliefs have to be mutually consistent, so it
must be possible, in principle at least, to integrate all of the true per-
spectives, even if only by conjoining them. At the end of the day, repre-
sentation is still perspectival, but perhaps the only perspective to which it
is intrinsically relative is our point of view as rational inquirers, or cogni-
tive systems, or (as Moore sometimes puts it) as processors of knowledge.
Maybe you can't know what things are like in themselves, but you can,
78 Chapter 7
at least in principle, know what they are like insofar as they are possible
objects of knowledge. If this sounds a bit truistic, so much the better.
That's encouraging as far as it goes, but it doesn't really go far. For one
thing (a point Moore doesn't emphasize), the perspectival relativists he
is trying to split the difference with are unlikely to grant him his ''basic
assumption" of a ready-made world. To the contrary, it's part and parcel
of the Kuhnian sort of view that thinkers with different paradigms ipso
facto live in different worlds. That's the metaphysical obverse of the
semantic claim that you can't integrate divergent paradigms. (Since Kuhn
takes content to be paradigm-relative, what you get if you conjoin repre-
sentations from different paradigms is just equivocations.) Indeed, the
many-worlds story is, arguably, just the incommensurable-paradigms story
translated from semantics to metaphysics; so, arguably, you can't have
one without the other. If it's not clear how literally we're to take either, or
even what taking them literally would amount to, that just shows what a
mess we're in.
And second, a point to which Moore is entirely alert, transcendental
idealism looks to be self-refuting. If you really can't say anything about
the world except as it is represented, then one of the things that you can't
say is that you can't say anything about the world except as it is repre-
sented. For the intended contrast is between how the world is as repre-
sented and how it is sans phrase. But how the world is sans phrase is how it
is not from a perspective, and it's part of the story we're trying to tell that
representation is perspectival intrinsically. It's worth remarking how much
worse off Moore is than Kant in this respect. Kant thought that there are
substantive constraints on the "form" of the possible objects of knowl-
edge, and hence that we can't know about what fails to conform to these
constraints. Moore, following the early Wittgenstein, thinks that there are
substantive constraints on the form of the possible objects of representa-
tion, and hence that we can't even think about what fails to conform to
these constraints. For Kant, transcendental idealism was strictly mere
metaphysics (that is to say, its truth is unknowable); for Moore it is
strictly mere babble (that is to say, its content is unspeakable). It is, I
think, a general rule that whenever philosophy takes the '1inguistic turn"
the mess it's in gets worse.
Transcendental idealism is nonsense by its own standards. So be it:
"[T]his leaves scope for all sorts of distinctions. . . . We can think of the
Nonsense Constraint as offering the following guideline when it comes to
making sense of t h ~ schema 'A is shown that x: namely to prescind alto-
gether from considerations of content and to think more in aesthetic
terms .... To say of some piece of nonsense that it is the result of attempt-
ing to express the inexpressible is something like [sic] making an aesthetic
evaluation. It concerns what might be called, justly, if a little grandilo-
Review of Moore's Points of View 79
quently, the music of words" (202). One is powerfully reminded of Frank
Ramsey's salubrious wisecrack about the embarrassing bits of the Trac-
tatus: 'What can't be said can't be said and it can't be whistled either."
I'll spare you the details, which are complicated and, in my view, much
less than convincing. Roughly, we have lots of "inexpressible" knowl-
edge. Its being inexpressible is somehow connected with its being know-
ing how (e.g., knowing how to process knowledge) rather than knowing
that, though I find this connection obscure. Anyhow, sometimes we're
driven to try to express the inexpressible (that we are has something to
do with our aspiring to be infinite); when we do so, we talk nonsense.
For example, we say things such as that representations can (or can't) be
absolute. Since this is nonsense, saying it can, of course, communicate
nothing. But that we are inclined to say it shows all sorts of things. For
example,. the nonsense we're tempted to talk about value shows that
"things are not of value tout court. Nothing is. However, another thing
we are shown is that they are of value tout court. Our aspiration to be
infinite, precisely in determining that these things are of value for us,
leads to our being shown this" (xiii). It's tempting to dismiss this sort of
thing as merely irritating. But if you propose to do so, you have to figure
out how to avoid it while nevertheless relativizing meaning to per-
spective and retaining a respectable notion of objectivity. It is enormously
to Moore's credit that he has faced this squarely.
But, after all, where does having faced it get him? According to enemies
of objectivity like Kuhn, it's false to saythat the world chooses between
our theories. According to friends of objectivity like Moore, it's nonsense
to say that the world chooses between our theories. Objectivity might
reasonably complain that with such friends it doesn't need existentialists.
What's remarkable is that neither side of the argument considers, even for
a moment, abandoning the premise that's causing the troubles; namely,
that content is ipso facto perspectival and holistic.
Perspectival theories of meaning make objectivity look terribly hard. So
much the worse for perspectival theories of meaning; objectivity is easy.
Here's some: My cat has long whiskers. He doesn't (as Moore would be
the first to agree) have long whiskers "from a perspective" or "relative to
a conceptual syste:n" or "imminently"; he has long whiskers tout court.
That's what my cat is like even when I am not there (unless, however,
someone should snip his whiskers while I'm not looking; which is not the
sort of thing that Moore is worrying about). Prima facie, the ontological
apparatus that's required for my cat to have long whiskers is extremely
exiguous; all that's needed is my cat, long whiskers, and the state of affairs
that consists of the former having the latter.
Likewise, it is possible for me to represent my cat' s having long
whiskers. Indeed, I just did. The semantic apparatus required for this is
80 Chapter 7
also exiguous, and for much the same reasons. All I need is to be able to
represent my cat, long whiskers, and the state of affairs that consists of
the former having the latter. First blush at least, neither my eat's having
long whiskers nor my representing it as having them would seem to
imply a perspective or an outlook or a desire for in6nity, or anything of
the sort; the objectivity of the fact is about all that the objectivity of the
representation requires. So whence, exactly, all the angst?
That all this is so, and objectively so, is far more obvious than any
general principles about the relativity of content to perspectives or con-
ceptual systems. If, therefore, your perspedivist semantics leads you to
doubt that it is objectively so, or that it can coherently be said to be
objectively so, it's the semantics and not the cat that you should consider
getting rid of. That this methodological truism should have escaped so
many very sophisticated philosophers, Moore among them, seems to me
to be among the wonders of the age.
Part III
Cognitive Architecture
Chapter 8
Review of Paul Churchland's The Engine of Reason,
The Seat of the Soul
[W]e are now in a position to explain how our vivid sensory experience arises in
the sensory cortex of our brains ... [and is] embodied in a vast chorus of neural
activity ... to explain how the [nervous system and the musculature] perform
the cheetah's dash, the falcon's strike, or the ballerina's dying swan . ... [W]e can
now understand how the infant brain slowly develops a framework of concepts
... and how the matured brain deploys that framework almost instantaneously:
to recognize similarities, to grasp analogies, and to anticipate both the immediate
and the distant future.
-Paul Churchland, The Engine of Reason, The Seat of the Soul, pp. 4-7
I do think that is naughty of Professor Churchland. For one thing, none of
it is true. We don't know how the brain deploys its concepts to achieve
perception and thought; or how it develops them; or even what concepts
are. We don't know how the motor system contrives the integration of
the lips, tongue, lungs, velum, and vocal cords in the routine utterance of
speech, to say nothing of special effects like the approximation of dying
swans by whole ballerinas. Nor do we know, even to a first glimmer, how
a brain (or anything else that is physical) could manage to be a locus of
conscious experience. This last is, surely, among the ultimate metaphy-
sical mysteries; don't bet on anybody ever solving it.
Nor does Churchland's hyperoptimistic account of the current situation
in "cognitive neuroscience" represent anything like the consensus view
among practicioners. I'm sure this is so, because I've lately made a little
game of reading the above and related quotations to friends in linguistics,
cognitive psychology, physiological psychology, philosophy, and com-
puter theory. I had hoped to pass along some of their comments for your
amusement, but those who weren't dumbstruck used words that even The
New Yorker still won't print.
What makes this naughty of Churchland, and not just ill advised, is that
his book purports to be a popular introduction. Churchland wants to let
the layman in on what's been going on in mind/brain studies, and on
84 Chapter 8
implications and applications all the way from science and technology
to morality and jurisprudence. For example, not only do "[a]rtificial neural
networks have the power to change the way we do science" (314) but
"[the] new scanning techniques will allow us to accumulate a large data-
base concerning individual profiles of brain function, from normals to
violent sociopaths .... A large number of such pairs will constitute a train-
ing set for the sort of diagnostic network we desire, a network that, once
trained, will accurately diagnose certain types of brain dysfunction and
accurately predict problematic social behavior" (311).
If you're in the research business, you will recognize at once the rhet-
oric of technohype. It is the idiom of grant proposals and of interviews in
the Tuesday New York Science Times: The breakthrough is at hand; this time
we've got it right; theory and practice will be forever altered; we have really
made fantastic progress, and there is now general agreement on the basics; fur-
ther funding is required. Professionals are inured to this sort of stuff and
write it, and write it off, as a matter of course. I hope that the laity will
prove to be equally cynical. For Churchland' s book is, in fact, not an
overview but a potboiler. It's a polemic, generally informed and readable,
for a certain quite parochial account of the methodology and substance of
psychology: what they call in the trade 'West coast" cognitive science.
(Perhaps it goes without saying that 'West coast," in this usage, denotes a
spiritual rather than a geographical condition.)
Methodologically, Churchland endorses a research strategy for the
study of cognition that is rooted in neurological modeling. This will seem
to many like simply common sense. It's the brain that makes intelligence
and consciousness; so if you want to know what intelligence and con-
sciousness eire, it's brains that you need to look at. Well, maybe so, but
also maybe not. Compare: '1f it's flight that you want to understand, what
you need to look at is feathers." Actually, it's in the nature of research
strategies that they get vindicated after the fact if at all. One way to
study the mind/brain is to try to understand its chemistry and physiol-
ogy. Another way is to try to simulate its characteristic behavioral effects.
Another way is to forget about its behavioral effects and try to under-
stand its cognitive competence. Another way is to construct experimental
environments in which the details of its information processing can be
selectively scrutinized. Another way is to examine its evolutionary pre-
cursors; or the character of its ecological adaptivity; or the course of
its ontogeny, and so on; and there are surely lots of ways to study the
mind/brain that no one's thought of yet. In fact, none of the strategies
that we've tried so far has been extraordinarily successful, .the ones that
Churchland likes included. Probably the right thing to do is just what
we are doing: work on all of them together and hope that something will
tum up.
Review of Churchland' s The Engine of Reason 85
Methodology aside, the actual theory of the mind/brain that Church-
land is pushing is a version of connectionism according to which per-
ception and thought are carried out by "parallel distributed processors"
that are organized into "neural networks." And his message is that con-
nectionism is inevitable, so you had better learn to like it: ''You came to
this book assuming that the basic units of human cognition are states such
as thoughts, beliefs, perceptions, desires, and preferences .... [By contrast]
... it is now modestly plain that the basic unit of computation is the acti-
vation vector. It is now fairly clear that the basic unit of computation is the
vector-to-vector transformation. And it is now evident that the basic unit of
memory is the synaptic weight configuration [p. 322] .... It is now beyond
serious doubt that this is the principal form of computational activity in
all biological brains" (266). So, like, get with it.
Let us, nonetheless, do our best to keep our calm. To begin with,
"neural network" is a name (not, in the first instance, for a network of
neurons, but) for a certain kind of computer, the organization of which
may, or may not, throw some light on some aspects of human cognition.
It's a long story how these devices work; suffice it that they differ, in
various ways, from the "classical" architecture of Turing machines, or von
Neumann machines, or desktop computers. One of the differences is that
networks don't exhibit the distinction between program and memory that
is characteristic of the more familiar computational devices. Rather, both
the current computational proclivities of a network and the residual effects
of its computational history are encoded by connections between a very
large number of relatively simple, switch-like elements.
Networks compute in parallel, by sending waves of activation through
the connected elements. How much of the activation goes where depends
on which connections are operative, and which connections are operative
depends on the network's previous activations. You can make this picture
look like the cutting edge of science if you think of the elements as
neurons interactivating one another at synapses of variable resistance.
But you can equally make it look quite dreary and recidivist: Think of the
elements as "ideas" and call the connections between them "associations,"
and you've got a psychology that seems no great advance on David
Hume. I'm sorry if you find that disappointing, but there is worse to come.
Since their computational histories affect their computational procliv-
ities, networks are able to learn. Much has been made of this, but in fact
it's a tautology, not a breakthrough. Any device whose computational
proclivities are labile to its computational history is thereby able to learn,
and this truism includes networks and classical machines indiscriminately.
What distinguishes between them is that, although both can learn, the
former can't be programmed but have to be trained. As it turns out, that's
a mixed blessing.
86 Chapter 8
Because computation in networks is massively parallel (in effect, the
elements are interactivated simultaneously) they can sometimes solve
problems very fast; that's the good news. But which computations a net-
work is disposed to perform depends on the character of the connections
between many, many elements. To set the strength of each of these con-
nections "by hand" would be astronomically time consuming even if you
know which settings would produce the computations that you want-
which typically you don't and which there is no general procedure for
finding out. That's the bad news. The discovery that parallel machines are
likely to be fast but hard to program (hence, by the way, hard to debug if
something goes wrong) is also not a breakthrough; it's a commonplace of
the computer scientist's art. 'West coast" cognitive science is betting that
the benefits of massively parallel computing are probably worth the costs;
''East coast" cognitive science is betting that they're probably not. In any
event, networks are just one of the options; there are as many kinds of
computational architectures as you like that are intermediates between
purely serial and purely parallel machines. No doubt one of them (quite
possibly not one that's now even imaginable) will turn out to be right for
modeling the computations that the brain performs. That's all that any-
body's got so far.
If that's all that anybody' s got so far, then what is the fuss about7 I
mean, how do you promote that sort of situation into a scientific break-
through of the first magnitude? Here's how. You hew to the main maxim
of technohype: What doesn't work, you don't mention. Some illustrations
from Churchland follow.
Classical machines '1eam" inductive generalizations by executing soft-
ware programs that extract statistical regularities from their input data.
Networks '1eam" inductive generalizations because their training alters
their internal connections in ways that model such statistical regularities.
(It typically requires a lot-I mean, a lot-of training to produce such
alterations. That's the price that networks pay for massively parallel com-
putation.) In either case, you end up with a representation of the statistical
regularities in the input; so far, there is no evidence at all that the two kinds
of machines differ in what such generalizations they can learn. Churchland
is deeply impressed by how much better than people trained networks are
at certain kinds of perceptual discrimination tasks (at distinguishing sonar
pictures of mines from sonar pictures of rocks, for example). But that's the
wrong comparison; all it shows is that people aren't very good at doing
statistics in their heads. The right comparison is how networks fare in
competition with conventional computers that are programmed with sta-
tistical analysis software. When you make such comparisons, it turns out
(in any of the several cases I've heard of) that the two kinds of machines
Review of Churchland's The Engine of Reason 87
succeed in much the same tasks, but the networks typically are markedly
slower. Churchland reports none of these results.
The worries that Churchland doesn't talk about go still deeper. If you
want to ask whether the correlation of property A with property B is
statistically reliable in a body of data, you have to precode the data for
A-ness and B-ness. You can't assess the inductive likelihood that the next
red thing that you come across will be a square unless you've got your
previous experience coded for redness and for squareness. Often enough,
this question of what to code for in the data is the hardest part of solving
an inductive problem: The statistics of finding correlations is relatively
straightforward once you know which properties to look for correlations
between. Networks have this problem too, of course, and Churchland has
no more idea of how to solve it than I do; that is, no idea at all. What he
does instead is rename it, so what used to get called "properties" Church-
land calls "dimensions." This too is not a breakthrough.
Some of the things that minds respond to do seem to have a natural
taxonomy. Think of colors (one of Churchland's parade examples): You
can construct a space with two dimensions of hue (red to green and blue
to yellow) and one dimension of saturation (black to white), and each
color will occupy a unique position in this space. That is, you can identify
each color with the unique triple of numbers (a "vector'') that gives its
position on each of the dimensions. With such a scheme for coding the
data in hand, it's duck soup to answer statistical questions like whether, in
a certain sample of colors, the saturation is correlated with the hue.
Churchland is in love with dimensions, vectors, and prototypes. His
idea is to do for concepts at large what hue and saturation allow you to
do for colors: locate them in a "multidimensional vector space" and then
look for correlations between the vector values. So, for example, your
concept of a certain dog is maybe a vector of (proto )typical values along
whatever the dimensions are that define dog-space; and your concept of
dogness per se is maybe a vector of prototypical values along whatever
the dimensions are that define animal-space. And so on. (Strictly speaking,
vectors determine points in a vector space. But if, like Churchland, you
think that concepts are too vague and fuzzy to be points, you can take
them to be the regions around the points. Fuzzy is one of the things that
West coast cognitive science likes best.)
So concepts are regions in a space of cognitive dimensions; that's the
breakthrough. ''What cognitive dimensionsr' Well may you ask. You can
get experimental subjects to estimate how similar and how different <;olor
samples are; and, if you analyze their answers, you can recover the per-
ceptual dimensions of color. But colors are special in that respect; they
really are all perceived as occupying positions in a space of a (relatively)
few dimensions, which is to say that they are seen as similar and different
88 Chapter 8
in a (relatively) small number of respects. Surely, nothing of the sort is
true in general of the things that our concepts apply to? Surely there isn't a
dimensional space in which, as it might be, the color red, lawyers, Easter
Sunday, and the number three are variously propinquitous7 Red is more
similar to orange than it is to green. Is it also more similar to Easter
Sunday than it is to lawyers? If this question strikes you as a nonsense,
beware the guy who tells you that concepts are vectors in cognitive
space.
To be sure, it's always possible to contrive a space ad hoc; perhaps
that's why Churchland seems so confident that things will tum out well
for vector coding. For example, you could have a space one dimension of
which represents degrees of being Easter Sunday. Easter Sunday could then
be represented, by a vector in this space, as having a higher value on that
dimension than anything else. If you've got only finitely many concepts
to taxonomize, you will need only finitely many spatial dimensions to
specify such vectors. (And if you've got infinitely many concepts, spaces
of infinitely many dimensions are likewise available.) But, of course, this is
just an unedifying kind of fooling around; it's an imperspicuous way of
saying "what makes Easter Sunday different from lawyers is that Easter
Sunday is Easter Sunday and lawyers aren't." That is not news, and the
appeal to vectors, dimensions, and prototypes isn't doing any work.
Churchland writes as though he thinks that prototype vectors approx-
imate a general answer to the abysmally deep question of how a mind
represents the things that it thinks about. But the way he uses them, they
don't; they're just a way of talking. So, having described the progress of
mechanics from Aristotle to Einstein as the transformation of their respec-
tive prototype vectors, Churchland remarks that such achievements are
continuous with "the very same activities of vector processing, recurrent
manipulation, prototype activation, and prototype evaluation as can be
found in some of the simplest ... cognitive activities ... [these being] the
mechanisms embodied in a large and highly trained recurrent neural net-
work" (121). The difference is just that Einstein's prototype "geodesic
paths in a non-Euclidean space-time--is admittedly arcane to most of us"
(120). That's how you say in technohypespeak that Einstein was smart
enough to figure out that planetary orbits are straight lines in space-time
but you and I probably wouldn't have been. As far as I can tell, that's all
it says.
Saying that we conceptualize things as regions in a multidimensional
space is saying no more than that we conceptualize things. The appear-
ance of progress is merely rhetorical. If you want vectors and dimensions
to be a breakthrough and not just loose talk, you've got to provide some
serious proposals about what the dimensions of cognitive space are. This is
exactly equivalent to asking for a revealing taxonomy of the kinds of
Review of Churchland' s The Engine of Reason 89
concepts that people can have. Short of the puzzles about consciousness,
there isn't any problem about the mind that's harder or less well under-
stood. It goes without saying (or anyhow it should go without saying)
that Churchland doesn't know how to provide such a taxonomy. Nor do
I. Nor does anybody else.
Oh well, we've all heard psychology technohype before. Last time it
was expert systems. The time before last it was operant conditioning and
reinforcement theory. The time before that it was classical conditioning.
The time before that it was the association of ideas. If you want a breath
of reality, just to dear your head, look up "parallel distributed processing"
in the fourth edition of Henry Gleitman's majesterial text Psychology.
POPs get about one page (out of eight hundred). Gleitman concludes:
''The PDP approach is very ambitious ... [but] as yet it is much too early
to come to any final evaluation. . . . [W]hatever we might say today will
almost surely be out of date tomorrow" (291-292). Quite so. That's the
way it goes with technohype.
The fact is (have I mentioned this before?) that we don't know how
minds work, though we're pretty good at explaining certain aspects of
perception. Nor do we know how brains work, if that's a different ques-
tion, though we're getting pretty good at finding out where it is that they
do their mental stuff. Churchland duly retails the latest gossip about MRis
and PETs and other such capital intensive hardware. This new technology
may turn out to be useful for psychology, or again it may not. If you
don't know what the carburetor is, you may not care all that much where
the .carburetor is. Ditto the neural locations of consciousness and thought.
Churchland says he wrote his book in part because "research into neu-
ral networks ... has produced the beginnings of a real understanding of
how the biological brain works, a real understanding, that is, of how you
work [sic] . ... [T]his idea may be found threatening, as if your innermost
secrets were about to be laid bare or made public" (3 ). Maybe that is what
they're feeling threatened by in San Diego, where the sun always shines
and even the poor are moderately rich. But here in New York the weather
is awful, the infrastructure is collapsing and even the moderately rich are
poor: We should also worry about neural nets?
Chapter 9
Connectionism and the Problem of Systematicity:
Why Smolensky's Solution Doesn't Work
Jerry Fodor and Brian McLaughlin
Introduction
In two recent papers, Paul Smolensky (1987, 1988b) responds to a chal-
lenge Jerry Fodor and Zenon Pylyshyn (Fodor and Pylyshyn, 1988) have
posed for connectionist theories of cognition: to explain the existence of
systematic relations between cognitive capacities without assuming that
cognitive processes are causally sensitive to the constituent structure of
mental representations. This challenge implies a dilemma: if connection-
ism can't account for systematicity, it thereby fails to provide an adequate
basis for a theory of cognition; but if its account of systematicity requires
mental processes that are sensitive to the constituent structure of mental
representations, then the theory of cognition it offers will be, at best, an
implementation architecture for a "classical" (language of thought) model.
Smolensky thinks connectionists can steer between the horns of this
dilemma if they avail themselves of certain kinds of distributed mental
representation. In what follows, we will examine this proposal.
Our discussion has three parts. In section I, we briefly outline the phe-
nomenon of systematicity and its classical explanation. As we will see,
Smolensky actually offers two alternatives to this classical treatment, cor-
responding to two ways in which complex mental representations can
be distributed; the first kind of distribution yields complex mental repre-
sentations with "weak compositional structure," the second yields com-
plex mental representations with "strong compositional structure." We
will consider these two notions of distribution in tum: in section II,
we argue that Smolensky' s proposal that complex mental representations
have weak compositional structure should be rejected as inadequate to
explain systematicity and also on internal grounds; in section III, we argue
that postulating mental representations with strong compositional struc-
ture also fails to provide for an explanation of systematicity. The upshot
will be that Smolensky avoids only one hom of the dilemma that Fodor
and Pylyshyn proposed. We shall see that his architecture is genuinely
nonclassical since the representations he postulates are not "distributed
over" constituents in the sense that classical representations are; and we
92 Chapter 9
shall see that for that very reason Smolensky' s architecture leaves sys-
tematicity unexplained.
I The Systematicity Problem and Its Classical Solution
The systematicity problem is that cognitive capacities come in clumps.
For example, it appears that there are families of semantically related
mental states such that, as a matter of psychological law, an organism is
able to be in one of the states belonging to the family only if it is able to
be in many of the others. Thus you don't find organisms that can learn to
prefer the green triangle to the red square but can't learn to prefer the red
triangle to the green square. You don't find organisms that can think the
thought that the girl loves John but can't think the thought that John
loves the girl. You don't And organisms that can infer P &om P & Q & R
but can't infer P &om P & Q. And so on over a very wide range of cases.
For the purposes of this paper, we assume without argument:
(i) that cognitive capacities are generally systematic in the above
sense, both in humans and in many in&ahuman organisms;
(ii) that it is nomologically necessary (hence counterfactual sup-
porting) that this is so;
(iii) that there must therefore be some psychological mechanism in
virtue of the functioning of which cognitive capacities are system-
atic; and
(iv) that an adequate theory of cognitive architecture should
exhibit this mechanism.
Any of (i)-(iv) may be viewed as tendentious; but, so far as we can tell,
all four are accepted by Smolensky. So we will take them to be common
ground in what follows.
The classical account of the mechanism of systematicity depends cru-
cially on the idea that mental representation is language-like. In particular,
mental representations have a combinatorial syntax and semantics. We
tum to a brief discussion of the classical picture of the syntax and seman-
tics of mental representations; this provides the basis for understanding
the classical treatment of systematicity.
1
Classical Syntax and Classical Constituents
The Classical view holds that the syntax of mental representations is like
the syntax of natural language sentences in the following respect: both
include complex symbols (bracketing trees) which are constructed out of
what we will call classical constituents. Thus, for example, the English
Connectionism and Systematicity 93
sentence "John loves the girl" is a complex symbol whose decomposition
into classical constituents is exhibited by some such bracketing tree as:
Sentence
~
~ J ~
John loves the girl
Correspondingly, it is assumed that the mental representation that is
entertained when one thinks the thought that John loves the girl is a
complex symbol of which the classical constituents include representa-
tions of John, the girl, and loving.
It will become dear in section III that it is a major issue whether the
sort of complex mental representations that are postulated in Smolensky' s
theory have constituent structure. We do not wish to see this issue
degenerate into a terminological wrangle. We therefore stipulate that, for
a pair of expression types EI and E2, the 6rst is a classical constituent of
the second only if the 6rst is tokened whenever the second is tokened. For
example, the English word "John" is a classical constituent of the English
sentence "John loves the girl," so every tokening of the latter implies a
tokening of the former (speci6cally, every token of the latter contains a
token of the former; you can't say "John loves the girl" without saying
"John").
2
Likewise, it is assumed that a Mentalese symbol which names
John is a classical constituent of the Mentalese symbol that means that
John loves the girl. So again a tokening of the one symbol requires a
tokening of the other. It is precisely because classical constituents have
this property that they are always accessible to operations that are defined
over the complex symbols that contain them; in particular, it is precisely
because classical mental representations have classical constituents that
they provide domains for structure-sensitive mental processes. We shall
see presently that what Smolensky offers as the "constituents" of con-
nectionist mental representations are nonclassical in this respect, and that
that is why his theory provides no account of systematicity.
Classical Semantics
It is part of the classical picture, both for mental representation and for
representation in natural languages, that generally when a complex
formula (e.g., a sentence) S expresses the proposition P, S's constituents
express (or refer to) the elements of P.
3
For example, the proposition that
94 Chapter 9
John loves the girl contains as its elements the individuals John and the
girl, and the two-place relation loving. Correspondingly, the formula
"John loves the girl," which English uses to express this proposition, con-
tains as constituents the expressions ''John," '1oves," and "the girl." The
sentence ''John left and the girl wept," whose constituents include the
formulas "John left" and "the girl wept," expresses the proposition that
John left and the girl wept, whose elements include the proposition that
John left and the proposition that the girl wept. And so on.
These assumptions about the syntax and semantics of mental repre-
sentations are summarized by condition (C):
(C) If a proposition P can be expressed in a system of mental repre-
sentation M, then M contains some complex mental representation
(a "mental sentence") S, such that S expresses P and the (classical)
constituents of S express (or refer to) the elements of P.
Systematicity
The classical explanation of systematicity assumes that (C) holds by
nomological necessity; it expresses a psychological law that subsumes all
systematic minds. It should be clear why systematicity is readily expli-
cable on the assumptiofl:S, first, that mental representations satisfy (C),
and, second, that mentalprocesses have access to the constituent structure
of mental representations. Thus, for example, since (C) implies that any-
one who can represent a proposition can ipso facto represent its elements,
it implies, in particular, that anyone who can represent the proposition
that John loves the girl can ipso facto represent John, the girl, and the
two-place relation loving. Notice, however, that the proposition that the
girl loves John is also constituted by these same individuals and relations.
So, then, assuming that the processes that integrate the mental repre-
sentations that express propositions have access to their constituents, it
follows that anyone who can represent John's loving the girl can also
represent the girl's loving John. Similarly, suppose that the constituents of
the mental representation that gets tokened when one thinks that P & Q
& R and the constituents of the mental representation that gets tokened
when one thinks that P & Q both include the mental representation that
gets tokened when one thinks that P. And suppose that the mental pro-
cesses that mediate the drawing of inferences have access to the con-
stituent structure of mental representations. Then it should be no surprise
that anyone who can infer P from P & Q & R can likewise infer P from
P&Q.
To summarize: the classical solution to the systematicity problem
entails that (i) systems of mental representation satisfy (C) (a fortiori,
complex mental representations have classical constituents); and (ii) mental
Connectionism and Systematicity 95
processes are sensitive to the constituent structure of mental representa-
tions. We can now say quite succinctly what our claim against Smolensky
will be: on the one hand, the cognitive architecture he endorses does not
provide for mental representations with classical constituents; on the
other hand, he provides no suggestion as to how mental processes could
be structure-sensitive unless mental representations have classical con-
stituents; and, on the third hand (as it were) he provides no suggestion as
to how minds could be systematic if mental processes aren't structure-
sensitive. So his reply to Fodor and Pylyshyn fails. Most of the rest of the
chapter will be devoted to making this analysis stick.
II Weak Compositionality
Smolensky' s views about "weak" compositional structure are largely
inexplicit and must be extrapolated from his "coffee story," which he tells
in both of the papers under discussion (and also in 1988a). We turn now
to considering this story.
Smolensky begins by asking how we are to understand the relation
between the mental representation COFFEE and the mental representation
CUP WITH COFFEE.
4
His answer to this question has four aspects that
are of present interest:
(i) COFFEE and CUP WITH COFFEE are activity vectors (accord-
ing to Smolensky' s weak compositional account, this is true of the
mental representations corresponding to all commonsense concepts;
whether it also holds for, e.g., technical concepts won't matter for
what follows). A vector is, of course, a magnitude with a certain
direction. A pattern of activity over a group of "units" is a state
consisting of the members of the group each having an activation
value of 1 or 0.
5
Activity vectors are representations of such pat-
terns of activity.
(ii) CUP WITH COFFEE representations contain COFFEE repre-
sentations as (nonclassical) constituents in the following sense: they
contain them as component vectors. By stipulation, a is a compo-
nent vector of b if there is a vector z such that a + z = b (where "+"
is the operation of vector addition). More generally, according to
Smolensky, the relation between vectors and their nonclassical con-
stituents is that the former are derivable from the latter by opera-
tions of vector analysis.
6
(iii) COFFEE representations and CUP WITH COFFEE representa-
tions are activity vectors over units which represent microfeatures
(units like BROWN, LIQUID, MADE OF PORCELAIN, etc.).
96 Chapter 9
(iv) COFFEE (and, presumably, any other representation vector)
is context-dependent. In particular, the activity vector that is the
COFFEE representation in CUP WITH COFFEE is distinct from the
activity vector that is the COFFEE representation in, as it might be,
GLASS WITH COFFEE or CAN WITH COFFEE. Presumably this
means that the vector in question, with no context specified, does
not give necessary conditions for being coffee. (We shall see later
that Smolensky apparently holds that it doesn't specify sufficient
conditions for being coffee either.)
Claims (i) and (ii) introduce the ideas that mental representations are
activity vectors and that they have (nonclassical) constituents. These ideas
are neutral with respect to the distinction between strong and weak com-
positionality so we propose to postpone discussing them until section III.
Claim (iii), is, in our view, a red herring. The idea that there are micro-
features is orthogonal both to the question of systematicity and to the
issue of compositionality. We therefore propose to discuss it only briefly.
It is claim (iv) that distinguishes the strong from the weak notion of com-
positional structure: a representation has weak compositional structure iff
it contains context-dependent constituents. We propose to take up the
question of context-dependent representation here.
We commence by reciting the coffee story (in a slightly condensed form).
Since, following Smolensky, we are assuming heuristically that units
have bivalent activity levels, vectors can be represented by ordered sets
of zeros (indicating that a unit is "off") and ones (indicating that a unit is
"on"). Thus, Smolensky says, the CUP WITH COFFEE representation
might be the following activity vector over microfeatures:
I-UPRIGHT CONTAINER
I-HOT LIQUID
0-GLASS CONTACTING WOOD
7
I-PORCELAIN CURVED SURFACE
I-BURNT ODOR
I-BROWN LIQUID CONTACTING PORCELAIN
I-PORCELAIN CURVED SURFACE
0-0BLONG SILVER OBJECT
I-FINGER-SIZED HANDLE
I-BROWN LIQUID WITH CURVED SIDES AND BOTTOM
8
This vector, according to Smolensky, contains a COFFEE representation
as a constituent. This constituent can, he claims, be derived from CUP
WITH COFFEE by subtracting CUP WITHOUT COFFEE from CUP
WITH COFFEE. The vector that is the remainder of this subtraction will
be COFFEE.
Connectionism and Systematicity 97
The reader will object that this treatment presupposes that CUP
WITHOUT COFFEE is a constituent of CUP WITH COFFEE. Quite so.
Smolensky is explicit in claiming that lithe pattern or vector representing
cup with coffee is composed of a vector that can be identified as a partic-
ular distributed representation of cup without coffee with a representation
with the content coffee" (I988b, IO).
One is inclined to think that this must surely be wrong. If you combine
a representation with the content cup without coffee with a representation
with the content coffee, you get not a representation with the content cup
with coffee but rather a representation with the self-contradictory content
cup without coffee with coffee. Smolensky' s subtraction procedure appears to
confuse the representation of cup without coffee (viz., CUP WITHOUT
COFFEE) with the representation of cup without the representation of
coffee (viz., CUP). CUP WITHOUT COFFEE expresses the content cup
without coffee; CUP combines consistently with COFFEE. But nothing
does both.
On the other hand, it must be remembered that Smolensky' s mental
representations are advertised as context dependent, hence noncomposi-
tional. Indeed, we are given no clue at all about what sorts of relations
his theory acknowledges between the semantic properties of complex
symbols and the semantic properties of their constituents. Perhaps in a
semantics where constituents don't contribute their contents to the sym-
bols they belong to, it's alright after all if CUP WITH COFFEE has CUP
WITHOUT COFFEE (or, for that matter, PRIME NUMBER, or GRAND-
MOTHER, or FLYING SAUCER, or THE LAST OF THE MOHICANS)
among its constituents.
In any event, to complete the story, Smolensky gives the following
features for CUP WITHOUT COFFEE:
I-UPRIGHT CONTAINER
0-HOT LIQUID
0-GLASS CONTACTING WOOD
I-PORCELAIN CURVED SURFACE
0-BURNT ODOR
0-BROWN LIQUID CONTACTING PORCELAIN
I-PORCELAIN CURVED SURFACE
0-0BLONG SILVER OBJECT
I-FINGER-SIZED HANDLE
0-BROWN LIQUID WITH CURVED SIDES AND BOTTOM, etc.
Subtracting this vector from CUP WITH COFFEE, we get the following
COFFEE representation:
98 Chapter 9
0-UPRIGHT CONTAINER
I-HOT LIQUID
0-GLASS CONTACTING WOOD
0-PORCELAIN CURVED SURFACE
I-BURNT ODOR
I-BROWN LIQUID CONTACTING PORCELAIN
0-PORCELAIN CURVED SURFACE
0-0BLONG SILVER OBJECT
0-FINGER-SIZED HANDLE
I-BROWN LIQUID WITH CURVED SIDES AND BOTTOM
That, then, is Smolensky' s "coffee story."
Comments
Microfeatures
It's common ground in this discussion that the explanation of system-
aticity must somehow appeal to relations between complex mental repre-
sentations and their constituents (on Smolensky's view, to combinatorial
relations between vectors). The issue about whether there are micro-
features is entirely orthogonal; it concerns only the question of which
properties are expressed by the activation states of individual units. (To
put it in more classical terms, it concerns the question of which symbols
constitute the primitive vocabulary of the system of mental representa-
tions.) If there are microfeatures, then the activation states of individual
units are constrained to express only (as it might be) "sensory" properties
(I987, I46). If there aren't, then activation states of individual units can
express not only such properties as being brown and being hot, but also
such properties as being coffee. It should be evident upon even casual
reflection that, whichever way this issue is settled, the constituency ques-
tion (viz., the question of how the representation COFFEE relates to the
rep resentation CUP WITH COFFEE remains wide open. We therefore
propose to drop the discussion of microfeatures in what follows.
Context-dependent Representation
As far as we can tell, Smolensky holds that the representation of coffee
that he derives by subtraction from CUP WITH COFFEE is context
dependent in the sense that it need bear no more than a "family resem-
blance" to the vector that represents coffee in CAN WITH COFFEE,
GLASS WITH COFFEE, etc. There is thus no single vector that counts as
the COFFEE representation, hence no single vector that is a component of
all the representations which, in a classical system, would have COFFEE
as a classical constituent.
Connectionism and Systematicity 99
Smolensky himself apparently agrees that this is the wrong sort of con-
stituency to account for systematicity and related phenomena. As he
remarks, "a true constituent can move around and fill any of a number of
different roles in different structures" (1988b, 11), and the connection
between constituency and systematicity would appear to tum on this. For
example, the solution to the systematicity problem mooted in section I
depends exactly on the assumption that tokens of the representation type
JOHN express the same content in the context ... LOVES THE GIRL as
they do in the context THE GIRL LOVES ... (viz., that they pick out John,
who is an element both of the proposition John loves the girl and of the
proposition the girl loves John). It thus appears, prima facie, that the ex-
planation of systematicity requires context-independent constituents.
How, then, does the assumption that mental representations have weak
compositional structure (that is, that mental representation is context
dependent) bear on the explanation of systematicity7 Smolensky simply
doesn't say. And we don't have a clue. In fact, having introduced the
notion of weak compositional structure, Smolensky to all intents and pur-
poses drops it in favor of the notion of strong compositional structure,
and the discussion of systematicity is carried out entirely in terms of the
latter. What he takes the relation between weak and strong compositional
structure to be, and for that matter, which kind of structure he thinks that
mental representations actually have is thoroughly unclear.
9
In fact, quite independent of its bearing on systematicity, the notion
of weak compositional structure as Smolensky presents it is of dubious
coherence. We close this section with a remark or two about this point.
It looks as though Smolensky holds that the COFFEE vector that you
get by subtraction from CUP WITH COFFEE is not a COFFEE repre-
sentation when it stands alone. ''This representation is indeed a repre-
sentation of coffee, but [only?] in a very particular context: the context
provided by cup [i.e., CUP]" (1987, 147). If this is the view, it has bizarre
consequences. Take a liquid that h ~ s the properties specified by the
microfeatures that comprise COFFEE in isolation-but that isn't coffee.
Pour it into a cup, et voila! it becomes coffee by semantical magic.
Smolensky would clearly deny that the vector COFFEE that you get
from CUP WITH COFFEE gives necessary conditions for being coffee,
since you'd get a different COFFEE vector by subtraction from, say,
GLASS WITH COFFEE. And the passage just quoted suggests that he
thinks it doesn't give sufficient conditions either.
10
But, then, if the micro-
features associated with COFFEE are neither necessary nor sufficient for
being coffee the question arises what, according to this story, does make a
vector a COFFEE representation; when does a vector have the content
coffee?
100 Chapter 9
As far as we can tell, Smolensky holds that what makes the COFFEE
component of CUP WITH COFFEE a representation with the content
coffee is that it is distributed over units representing certain microfeatures
and that it figures as a component vector of a vector that is a CUP WITH
COFFEE representation. As remarked above, we are given no details at all
about this reverse compositionality according to which the embedding
vector determines the contents of its constituents; how it is supposed
to work isn't even discussed in Smolensky' s papers. But, in any event, a
regress threatens since the question now arises: if being a component of a
CUP OF COFFEE representation is required to make a vector a COFFEE
representation, what is required to make a vector a CUP OF COFFEE
representation? Well, presumably CUP OF COFFEE represents a cup of
coffee because it involves the microfeatures it does and because it is a
component of still another vector; perhaps one that is a THERE IS A CUP
OF COFFEE ON THE TABLE representation. Does this go on forever? If
it doesn't, then presumably there are some vectors that aren't constituents
of any others. But now, what determines their contents? Not the contents
of their constituents, because by assumption, Smolensky' s semantics isn't
compositional (CUP WITHOUT COFFEE is a constituent of CUP WITH
COFFEE, etc.). And not the vectors that they are constituents of, because
by assumption, there aren't any of those.
We think it is unclear whether Smolensky has a coherent story
about how a system of representations could have weak compositional
structure.
What, in light of all this, leads Smolensky to embrace his account of
weak compositionality? Here's one suggestion: perhaps Smolensky con-
fuses being a representation of a cup with coffee with being a CUP WITH
COFFEE representation. Espying some cup with coffee on a particular
occasion, in a particular context, one might come to be in a mental state
that represents it as having roughly the microfeatures that Smolensky
lists. That mental state would then be a representation of a cup with coffee
in this sense: there is a cup of coffee that it's a mental representation of.
But it wouldn't, of course, follow that it's a CUP WITH COFFEE repre-
sentation; and the mental representation of that cup with coffee might be
quite different from the mental representation of the cup with coffee that
you espied on some other occasion or in some other context. So which
mental representation a cup of coffee gets is context dependent, just as
Smolensky says. But that doesn't give Smolensky what he needs to make
mental representations themselves context dependent. In particular, from
the fact that cups with coffee get different representations in different
contexts, it patently doesn't follow that the mental symbol that represents
something as being a cup of coffee in one context might represent some-
thing as being something else (a giraffe say, or the last of the Mohicans)
Connectionism and Systematicity 101
in some other context. We doubt that anything will give Smolensky that,
since we know of no reason to suppose that it is true.
In short, it is natural to confuse the true but uninteresting thought that
how you mentally represent some coffee depends on the context, with the
much more tendentious thought that the mental representation COFFEE
is context dependent. Assuming that he is a victim of this confusion
makes sense of many of the puzzling things Smolensky says in the coffee
story. Notice, for example, that all the microfeatures in his examples
express more or less perceptual properties (cf. Smolensky's own remark
that his microfeatures yield a "nearly sensory level representation.")
Notice, too, the peculiarity that the microfeature "porcelain curved sur-
face" occurs twice in the vectors for CUP WITH COFFEE, COFFEE, CUP
WITHOUT COFFEE and the like. Presumably, what Smolensky has in
mind is that when you look at a cup, you see two curved surfaces, one
going off to the left and the other going off to the right.
Though we suspect this really is what's going on, we won't pursue this
interpretation further, since if it's correct, the coffee story is completely
irrelevant to the question of what kind of constituency relation a COFFEE
representation has to a CUP WITH COFFEE representation; and that,
remember, is the question that bears on the issues about systematicity.
III Strong Compositional Structure
So much, then, for "weak" compositional structure. Let us tum to Smo-
lensky's account of "strong" compositional structure. Smolensky says that
[a] true constituent can move around and fill any of a number of dif-
ferent roles in different structures. Can this be done with vectors
encoding distributed representations, and be done in a way that
doesn't amount to simply implementing symbolic syntactic con-
stituency? The purpose of this section is to describe research show-
ing that the answer is affirmative. (1988b, 11)
The idea that mental representations are activity vectors over units, and
the idea that some mental representations have other mental representa-
tions as components, is common to the treatment of both weak and
strong compositional structure. However, Smolensky' s discussion of the
latter idea differs in several respects &om his discussion of the former.
First, he explicitly supposes that units have continuous activation levels
between 0 and 1; second, he does not invoke the idea of microfeatures
when discussing strong compositional structure; third, he introduces a
new vector operation (multiplication) to the two previously mentioned
(addition and subtraction); fourth, and most important, on his view strong
compositional structure does not invoke-indeed, it would appear to be
102 Chapter 9
incompatible with-the notion that mental representations are, context
dependent. So strong compositional structure does not exhibit the inco-
herence of Smolensky' s theory of context-dependent representation.
We will proceed as follows. First we briefly present the notion of
strong compositional structure. Then we tum to criticism.
Smolensky explains the notion of strong compositional structure, in
part, by appeal to the ideas of a tensor product representation and a
superposition representation. To illustrate these ideas, consider how a
connectionist machine might represent four-letter English words. Words
can be decomposed into roles (viz., ordinal positions that letters can
occupy) and things that can fill these roles (viz., letters). Correspondingly,
the machine might contain activity vectors over units that represent the
relevant roles (i.e., over the role units) and activity vectors over units that
represent the fillers (i.e., over the filler units). Finally, it might contain
activity vectors over units that represent filled roles (i.e., letters in letter
positions); these are the binding units. The key idea is that the activity
vectors over the binding units might be tensor products of activity vec-
tors over the role units and the filler units. The representation of a word
would then be a superposition vector over the binding units; that is, a
vector that is arrived at by superimposing the tensor product vectors.
The two operations used here to derive complex vectors from compo-
nent vectors are vector multiplication (in the case of tensor product vec-
tors) and vector addition (in the case of superposition vectors). These are
iterative operations in the sense that activity vectors that result from the
multiplication of role vectors and filler vectors might themselves repre-
sent the fillers of roles in more complex structures. Thus a tensor product
that represents the word "John" as "J" in first position, "o" in second
position, etc., might itself be bound to the representation of a syntactical
function to indicate, for example, that '1ohn" has the role subject-of in
"John loves the girl." Such tensor product representations could them-
selves be superimposed over yet another group of binding units to yield
a superposition vector that represents the bracketing tree (John) (loves
(the girl)).
11
It is, in fact, unclear whether this sort of apparatus is adequate to repre-
sent all the semantically relevant syntactic relations that classical theories
express by using bracketing trees with classical constituents. (There are,
for example, problems about long-distance binding relations, as between
quantifiers and bound variables.) But we do not wish to press this point.
For present polemical purposes, we propose simply to assume that each
classical bracketing tree can be coded into a complex vector in such a
fashion that the constituents of the tree correspond in some regular way
to components of the vector.
Connectionism and Systematicity 103
But to assume this is not, of course, to grant that either tensor product
or superposition vectors have classical constituent structure. In particular,
from the assumptions that bracketing trees have classical constituents and
that bracketing trees can be coded by activity vectors, it does not follow
that activity vectors have classical constituents. On the contrary, a point
about which Smolensky is himself explicit is vital in this regard: the com-
ponents of a complex vector need not even correspond to patterns of
activity over units actually in the machine. As Smolensky puts it, the
activity states of the Aller and role units can be "imaginary" even though
the ultimate activity vectors-the ones that do not themselves serve as
Aller or role components of more complex structures-must be actual
activity patterns over units in the machine. Consider again our machine
for representing four-letter words: the superposition pattern that repre-
sents, say, the word "John" will be an activity vector actually realized in
the machine. However, the activity vector representing "J" will be merely
imaginary, as will the activity vector representing the 6rst letter position.
Similarly for the tensor product activity vector representing "J" in the 6rst
letter position. The only pattern of activity that will actually be tokened
in the machine is the superposition vector representing "John."
These considerations are of central importance for the following reason.
Smolensky' s main strategy is, in effect, to invite us to consider the com-
ponents of tensor product and superposition vectors to be analogous to
the classical constituents of a complex symbol, and hence to view them as
providing a means by which connectionist architectures can capture the
causal and semantic consequences of classical constituency in mental
representations. However, the components of tensor product and super-
position vectors differ from classical constituents in the following way:
when a complex classical symbol is tokened, its constituents are tokened.
When a tehsor product vector or superposition vector is tokened, its
components are not (except per accidens). The implication of this differ-
ence, from the point of view of the theory of mental processes, is that
whereas the classical constituents of a complex symbol are, ipso facto,
available to contribute to the causal consequences of its tokening-in
particular, they are available to provide domains for mental processes-
the components of tensor product and superposition vectors can have no
causal status as such. What is merely imaginary can't make things happen,
to put this point in a nutshell.
We will return presently to what these differences imply for the treat-
ment of the systematicity problem. There is, however, a preliminary issue
that needs to be discussed.
We have seen that the-components of tensor product and superposition
vectors, unlike classical constituents, are not, in general, tokened when-
ever the activity vector of which they are the components is tokened. It is
104 Chapter 9
worth emphasizing, in addition, the familiar point that there is in general
no unique decomposition of a tensor product or superposition vector
into components. Indeed, given that units are assumed to have continuous
levels of activation, there will be infinitely many decompositions of a
given activity vector. One might wonder, then, what sense there is in
talk of the decomposition of a mental representation into significant
constituents, given the notion of constituency that Smolensky' s theory
provides.
Smolensky replies to this point as follows. Cognitive systems will be
dynamical systems; there will be dynamic equations over the activation
values of individual units, and these will determine certain regularities
over activity vectors. Given the dynamical equations of the system, cer-
tain decompositions can be especially useful for "explaining and under-
standing" the system's behavior. In this sense, the dynamics of a system
may determine "normal modes" of decomposition into components. So,
for example, though a given superposition vector in principle can be
taken to be the sum of many different sets of vectors, it may yet tum out
that we get a small group of sets-even a unique set-when we decom-
pose in the direction of normal modes; and likewise for decomposing
tensor product vectors. The long and short is that it could, in principle,
tum out that given the (thus far undefined) normal modes of a dynamical
cognitive system, complex superposition vectors will have it in common
with classical complex symbols that they have a unique decomposition
into semantically significant parts. Of course, it also could tum out that
they don't, and no ground for optimism on this point has thus far been
supplied.
Having noted this problem, however, we propose simply to ignore it.
So here is where we now stand: by assumption (though quite possibly
contrary to fact), tensor product vectors and superposition vectors can
code constituent structure in a way that makes them adequate vehicles
for the expression of propositional content; and, by assumption (though
again quite possibly contrary to fact), the superposition vectors that cog-
nitive theories acknowledge have a unique decomposition into semanti-
cally interpretable tensor product vectors, which in tum have a unique
decomposition into semantically interpretable filler vectors and role
vectors; so it's determinate which proposition a given complex activity
vector represents.
Now, assuming all this, what about the systematicity problem?
The first point to make is this: if tensor product or superposition vector
representation solves the systematicity problem, the solution must be
quite different from the classical proposal sketched in section I. True,
tensor product vectors and superposition vectors "have constituents" in
some suitably extended sense: tensor product vectors have semantically
Connectionism and Systematicity 105
evaluable components, and superposition vectors are decomposable into
semantically evaluable tensor product vectors. But the classical solution to
the systematicity problem assumes that the constituents of mental repre-
sentations have causal roles; that they provide domains for mental pro-
cesses. The classical constituents of a complex symbol thus contribute to
determining the causal consequences of tokening that symbol, and it
seems clear that the "extended" constituents of a tensor product or a
superposition representation can't do that. On the contrary, the compo-
nents of a complex vector are typically not even tokened when the com-
plex vector itself is tokened; they are simply constituents into which the
complex vector could be resolved consonant with decomposition in the
direction of normal modes. But, to put it crudely, the fact that six could be
represented as "3 x 2" cannot, in and of itself, affect the causal processes
in a computer (or a brain) in which six is represented as "6.'' Merely
counterfactual representations have no causal consequences; only actually
tokened representations do.
Smolensky is, of course, sensitive to the question whether activity vec-
tors really do have constituent structure. He defends at length the claim
that he has not distorted the notion of constituency in saying that they
do. Part of this defense adverts to the role that tensor products and
superpositions play in physical theory:
The state of the atom, like the states of all systems in quantum
theory, is represented by a vector in an abstract vector space. Each
electron has an internal state (its "spin"); it also has a role it plays in
the atom as a whole: it occupies some "orbital," essentially a cloud
of probability for finding it at particular places in the atom. The
internal state of an electron is represented by a "spin vector"; the
orbital or role of the electron (part) in the atom (whole) is repre-
sented by another vector, which describes the probability cloud.
The vector representing the electron as situated in the atom is the
tensor product of the vector representing the internal state of the
electron and the vector representing its orbital. The atom as a whole
is represented by a vector that is the sum or superposition of vec-
tors, each of which represents a particular electron in its orbital ....
(1988b, 19-20)
"So," Smolensky adds, "someone who claims that the tensor product
representational scheme distorts the notion of constituency has some
explaining to do" (1988b, 20).
The physics lesson is greatly appreciated; but it is important to be clear
on just what it is supposed to show. It's not, at least for present purposes,
in doubt that tensor products can represent constituent structure. The rel-
evant question is whether tensor product representations have constituent
106 Chapter 9
structure; or, since we have agreed that they may be said to have con-
stituent structure "in an extended sense," it's whether they have the kind
of constituent structure to which causal processes can be sensitive, hence
the kind of constituent structure to which an explanation of systematicity
might appeal. But we have already seen the answer to this question: the
constituents of complex activity vectors typically aren't "there," so if the
causal consequences of tokening a complex vector are sensitive to its
constituent structure, that's a miracle.
12
We conclude that assuming that mental representations are activation
vectors does not allow Smolensky to endorse the classical solution of the
systematicity problem. And, indeed, we think Smolensky would grant this
since he admits up front that mental processes will not be causally sensi-
tive to the strong compositional structure of mental representations. That
is, he acknowledges that the constituents of complex mental representa-
tions play no causal role in determining what happens when the repre-
sentations get tokened: "Causal efficacy was not my goal in developing
the tensor product representation" (1988b, 21). What a r ~ causally effica-
cious according to connectionists are the activation values of individual
units; the dynamical equations that govern the evolution of the system
will be defined over these. It would thus appear that Smolensky must
have some nonclassical solution to the systematicity problem up his
sleeve, some solution that does not depend on assuming mental processes
that are causally sensitive to constituent structure. So then, after all this,
what is Smolensky' s solution to the systematicity problem?
Remarkably enough, Smolensky doesn't say. All he does say is that he
''hypothesizes ... that ... the systematic effects observed in the processing
of mental representations arise because the evolution of vectors can be (at
least partially and approximately) explained in terms of the evolution of
their components, even though the precise dynamical equations apply
[only] to the individual numbers comprising the vectors and [not] at the
level of [their] constituents-i.e. even though the constituents are not
causally efficacious" (1988b, 21).
It is left unclear how the constituents ("components") of complex vec-
tors are to explain their evolution (even partially and approximately)
when they are, by assumption, at best causally inert, and at worst merely
imaginary. In any event, what Smolensky clearly does think is causally re-
sponsible for the "evolution of vectors" (and hence for the systematicity
of cognition) are unspecified processes that affect the states of activation
of the individual units (the neuron analogs) out of which the vectors are
composed. So, then, as far as we can tell, the proposed connectionist ex-
planation of systematicity (and related features of cognition) comes down
to this: Smolensky ''hypothesizes" that systematicity is somehow a con-
sequence of underlying neural processes.
13
Needless to say, if that is
Connectionism and Systematicity 107
Smolensky' s theory, it is on the one hand certainly true, and on the other
hand not intimately dependent upon his long story about fillers, binders,
tensor products, superposition vectors, and the rest.
By way of rounding out the argument, we want to reply to a question
raised by an anonymous Cognition reviewer, who asks: "couldn't Smolen-
sky easily build in mechanisms to accomplish the matrix algebra oper-
ations that would make the necessary vector explicit (or better yet, from
his point of view, ... mechanisms that are sensitive to the imaginary com-
ponents without literally making them explicit in some string of units)r'
14
But this misses the point of the problem that systematicity poses for con-
nectionists, which is not to show that systematic cognitive capacities are
possible given the assumptions of a connectionist architecture, but to
explain how systematicity could be necessary-how it could be a law that
cognitive capacities are systematic-given those assumptions.
15
No doubt it is possible for Smolensky to wire a network so that it sup-
ports a vector that represents aRb if and only if it supports a vector that
represents bRa; and perhaps it is possible for him to do that without
making the imaginary units explicit (though there is, so far, no proposal
about how to ensure this for arbitrary a, R, and b). The trouble is that,
although the architecture permits this, it equally permits Smolensky to
wire a network so that it supports a vector that represents aRb if and
only if it supports a vector that represents zSq; or, for that matter, if
and only if it supports a vector that represents the last of the Mohicans.
The architecture would appear to be absolutely indifferent between these
options.
In contrast, as we keep saying, in the classical architecture, if you meet
the conditions for being able to represent aRb, you cannot but meet the con-
ditions for being able to represent bRa; the architecture won't let you do so
because (i) the representation of a, R, and b are constituents of the repre-
sentation of aRb, and (ii) you have to token the constituents of the
representations that you token, so classical constituents can't be just
imaginary. So then: it is built into the classical picture that you can't think
aRb unless you are able to think bRa, but the connectionist picture is
neutral on whether you can think aRb even if you can't think bRa. But it
is a law of nature that you can't think aRb if you can't think bRa. So
the classical picture explains systematicity and the connectionist picture
doesn't. So the classical picture wins.
Conclusion
At one point in his discussion, Smolensky makes some remarks that we
find quite revealing: he says that even in cases that are paradigms of clas-
sical architectures (LISP machines and the like), "we normally think of the
108 Chapter 9
'real' causes as physical and far below the symbolic level: ... " Hence, even
in classical machines, the sense in which operations at the symbol level
are real causes is just that "there is ... a complete and precise algorithmic
(temporal) story to tell about the states of the machine described" at that
level (1988b, 20). Smolensky, of course, denies that there is a "comparable
story at the symbolic level in the human cognitive architecture ... that is a
difference with the classical view that I have made much of. It may be that
a good way to characterize the difference is in terms of whether the constituents
in mental structure are causally efficacious in mental processing" (1988b, 20;
our emphasis).
We say that this comment is revealing because it suggests a diagnosis:
it would seem that Smolensky has succumbed to a sort of generalized
epiphenomenalism. The idea is that even classical constituents participate
in causal processes solely by virtue of their physical microstructure, so
that even on the classical story it's what happens at the neural level that
really counts. Though the evolution of vectors can perhaps be explained
in a predictively adequate sort of way by appeal to macroprocesses like
operations on constituents, still, if you want to know what's really going
on-if you want the causal explanation-you need to go down to
the "precise dynamical equations" that apply to activation states of units.
That intentional generalizations can only approximate these precise
dynamical equations is among Smolensky' s recurrent themes. By conflat-
ing the issue about "precision" with the issue about causal efficacy, Smo-
lensky makes it seem that to the extent that macrolevel generalizations
are imprecise, macrolevel processes are epiphenomenal.
It would need a philosophy lesson to say all of what's wrong with this
view. Suffice it for present purposes that his argument iterates in a way
that Smolensky ought to find embarrassing. No doubt, we do get greater
precision when we go from generalizations about operations on con-
stituents to generalizations about operations on units. But if that shows
that symbol-level processes aren't really causal, then it must be that unit-
level processes aren't really causal either. After all, we get still more
precision when we go down from unit-sensitive operations to molecule-
sensitive operations, and more precision yet when we go down from
molecule-sensitive operations to quark-sensitive operations. The moral is
not, however, that the causal laws of psychology should be stated in
terms of the behavior of quarks. Rather, the moral is that whether you
have a level of causal explanation is a question, not just of how much
precision you are able to achieve, but also of what generalizations you are
able to express. The price you pay for doing psychology at the level of
units is that you lose causal generalizations that symbol-level theories are
able to state. Smolensky' s problems with capturing the generalizations
about systematicity provide a graphic illustration of these truths.
Connectionism and Systematicity 109
It turns out, at any event, that there is a crucial caveat to Smolensky' s
repeated claim that connectionist mechanisms can reconstruct everything
that's interesting about the notion of constituency. Strictly speaking, he
claims only to reconstruct whatever is interesting about constituents
except their causes and effects. The explanation of systematicity turns on
the causal role of the constituents of mental representations and is there-
fore among the casualties. Hilary Putnam, back in the days when he was
still a metaphysical realist, used to tell a joke about a physicist who
actually managed to build a perpetual motion machine; all except for a
part that goes back and forth, back and forth, back and forth, forever.
Smolensky' s explanation of systematicity has very much the character of
this machine.
We conclude that Fodor and Pylyshyn's challenge to connectionism
has yet to be met. We still don't have even a suggestion of how to
account for systematicity within the assumptions of connectionist cogni-
tive architecture. I6
Notes
I. Since the two are often confused, we wish to emphasize that taking systematicity for
granted leaves the question of compositionality wide open. The systematicity of cogni-
tion consists in, for example, the fact that organisms that can think aRb can think bRa
and vice versa. Compositionality proposes a certain explanation of systematicity, viz.,
that the content of thoughts is determined in a uniform way by the content of the
context-independent concepts that are their constituents, and that the thought that
bRa is constituted by the same concepts as the thought that aRb. So the polemical situa-
tion is as follows. If you are a connectionist who accepts systematicity, then you must
argue either that systematicity can be explained without compositionality, or that con-
nectionist architecture accommodates compositional representation. So far as we can tell,
Smolensky vacillates between these options; what he calls "weak compositionality"
favors the former and what he calls "strong compositionality" favors the latter.
We emphasize this distinction between systematicity and compositionality in light of
some remarks by an anonymous Cognition reviewer: "By berating the [connectionist)
modelers for their inability to represent the common-sense [uncontextualized) notion of
'coffee' ... Fodor and Mclaughlin are missing a key point-the models are not supposed
to do so. If you buy the ... massive context-sensitivity ... that connectionists believe in."
Our strategy is not, however, to argue that there is something wrong with connection-
ism because it fails to offer an uncontextualized notion of mental (or, mutatis mutandis,
linguistic) representation. Our argument is that if connectionists assume that mental rep-
resentations are context sensitive, they will need to offer some explanation of system-
aticity that does not entail compositionality; and they do not have one.
We don't, therefore, offer direct arguments for context-insensitive concepts in what
follows; we are quite prepared that "coffee" should have a meaning only in context.
Only, we argue, if it does, then some noncompositional account of the systematicity of
coffee-thoughts will have to be provided.
2. Though we shall generally consider examples where complex symbols literally contain
their classical constituents, the present condition means to leave it open that symbols
may have classical constituents that are not among their (spatia-temporal) parts.
110 Chapter 9
(For example, so far as this condition is concerned. It might be that the classical con-
stituents of a symbol include the values of a "fetch" operation that takes the symbol as
an argument.)
J. We assume that the elements of propositions can include, for example, individuals,
properties, relations, and other propositions. Other metaphysical assumptions are, of
course, possible. For example, it is arguable that the constituents of propositions include
individual concepts (in the Fregean sense) rather than individuals themselves; and so on.
Fortunately, it is not necessary to enter into these abstruse issues to make the points that
are relevant to the systematidty problem. All we really need is that propositions have
internal structure, and that, characteristically, the internal structure of complex mental
representations corresponds in the appropriate way to the internal structure of the prop-
ositions that they express.
4. The following notational conventions will fadlitate the discussion: we will follow stan-
dard practice and use capitalized English words and sentences as canonical names for
mental representations. (Smolensky uses italidzed English expressions instead.) We stip-
ulate that the semantic value of a mental representation so named is the semantic value
of the corresponding English word or sentence, and we will italicize words or sentences
that denote semantic values. So, for example, COFFEE is a mental representation
that expresses (the property of) being coffee (as does the English word "coffee"); JOHN
LOVFS THE GIRL is a mental representation that expresses the proposition that John
loves tht girl; and so forth. It is important to notice that our notation allows that the
mental representation JOHN LOVFS THE GIRL can be atomic and the mental repre-
sentation COFFEE can be a complex symbol. That is, capitalized expressions should be
read as the names of mental representations rather than as their structural descriptions.
S. Smolensky apparently allows that units may have continuous levels of activation from
0 to 1. In telling the coffee story, however, he generally assumes bivalence for ease of
exposition.
6. As we shall see below, when an activity vector is tokened, its component vectors
typically are not. So the constituents of a complex vector are, ipso facto, nonclassical.
7. Notice that this microfeature is "off" in CUP WITH COFFEE, so it might be wondered
why Smolensky mentions it at all. The explanation may be this: operations of vector
combination apply only to vectors of the same dimensionality. In the context of the
weak constituency story, this means that you can only combine vectors that are activity
patterns over the same units. It follows that a component vector must contain the same
units (though possibly at different levels of activation) as the vectors with which it
combines. Thus if GRANNY combines with COFFEE to yield GRANNY'S COFFEE,
GRANNY must contain activation levels for all the units in COFFEE and vice versa.
In the present example, it may be that CUP WITH COFFEE is required to contain a
0-activation level for GLASS CONTACTING WOOD to accommodate cases where the
former is a component of some other vector. Similarly with OBLONG SILVER OBJECT
(below) since cups with coffee often have spoons in them.
&. Presumably Smolensky does not take this list to be exhaustive, but we don't know how
to continue it. Beyond the remark that although the microfeatures in his examples corre-
spond to "nearly sensory-level representation(sf' that fad is "not essential," Smolensky
provides no account at all of what determines which contents are expressed by micro-
features. The question thus arises why Smolensky assumes that COFFEE is not itself a
microfeature. In any event, Smolensky repeatedly warns the reader not to take his
examples of microfeatures very seriously, and we don't.
9. They can't have both; either the content of a representation is context dependent or it's
not. So, if Smolensky does think that you need strong compositional structure to explain
systematidty, and that weak compositional structure is the kind that connectionist
Connectionism and Systematicity 111
representations have, then it would seem that he thereby grants Fodor and Pylyshyn' s
claim that connectionist representations_ can't explain systematicity. We find this all very
mysterious.
10. If they were necessary and sufficient, COFFEE wouldn't be context dependent.
11. The function of the brackets in a classical bracketing tree is precisely to exhibit its
decomposition into constituents; and when the tree is well formed this decomposition
will be unique. Thus the bracketing of "(John) (loves) (the girl)" implies, for example,
both that "the girl" is a constituent and that '1oves the" is not.
12. It's a difference between psychology and physics that whereas psychology is about the
casual laws that govern tokenings of (mental) representations, physics is about the causal
laws that govern (not mental representations but) atoms, electrons, and the like. Since
being a representation isn't a property in the domain of physical theory, the question of
whether mental representations have constituent structure has no analogue in physics.
13. More precisely: we take Smolensky to be claiming that there is some property D, such
that if a dynamical system has D its behavior is systematic, and such that human be-
havior (for example) is caused by a dynamical system that has D. The trouble is that this
is a platitude, since it is untendentious that human behavior is systematic, that its causa-
tion by the nervous system is lawful and that the nervous system is dynamical. The
least that has to happen if we are to have a substantive connectionist account of system-
aticity is: first, it must be made clear what property D is, and second, it must be shown
that Dis a property that connectionist systems can have by law. Smolensky's theory
does nothing to meet either of these requirements.
14. Actually, Smolensky is forced to choose the second option. To choose the first would be,
in effect, to endorse the classical requirement that tokening a symbol implies tokening its
constituents-in which case, the question arises once again why such a network isn't
an implementation of a language-of-thought machine. Just as Smolensky mustn't allow
the representations of roles, fillers, and binding units to be subvectors of superposition
vectors if he is to avoid the "implementation" hom of Fodor and Pylyshyn' s dilemma,
so too he must avoid postulating mechani5ms that make role, filler, and binding units
explicit (specifically, accessible to mental operations) whenever the superposition vectors
are tokened. Otherwise he again has symbols with classical constituents and raises the
question of why the proposed device isn't a language-of-thought machine. Smolensky's
problem is that the very feature of his representations that makes them wrong for
explaining systematicity (viz., that their constituents are allowed to be imaginary) is the
one that they have to have to ensure that they aren't classical.
15. Fodor and Pylyshyn were very explicit about this. See, for example, 1988, 48.
16. Terence Horgan remarks (personal communication)," ... often there are two mathemati-
cally equivalent ways to calculate the time-evolution of a dynamical system. One is to
apply the relevant equations directly to the numbers that are elements of a single total
vector describing the initial state of the system. Another way is to mathematically
decompose that vector into component normal-mode vectors, then compute the time
evolution of each [of these] ... and then take the later state of the system to be described
by a vector that is the superposition of the resulting normal-mode vectors." Computa-
tions of the former sort are supposed to be the model for operations that are "sensitive"
to the components of a mental representation vector without recovering them. (Even in
the second case, it's the theorist who recovers them in the course of the computations by
which he makes his predictions. This does not, of course, imply that the constituents
thus "recovered" participate in causal processes in the system under analysis.)
Chapter 10
Connectionism and the Problem of Systematicity
(Continued): Why Smolensky's Sqlution Still Doesn't
Work
Paul Smolensky has recently announced that the problem of explaining the com-
positionality of concepts within a connectionist framework is solved in principle.
Mental representations are vectors over the activity states of connectionist
"units," but the vectors encode classical trees, whose structural properties in turn
"acausally" explain the facts of compositionality. This sounds suspiciously like
the offer of a free lunch, and it turns out, upon examination, that there is noth-
ing to it.
Human cognition exhibits a complex of closely related properties-
including systematicity, productivity, and compositionality-which a
theory of cognitive architecture ignores at its peril.
1
If you are stuck with
a theory that denies that cognition has these properties, you are dead and
gone. If you are stuck with a theory that is compatible with cognition
having these properties but is unable to explain why it has them, you are,
though arguably still breathing, clearly in deep trouble. There are, to be
sure, cognitive scientists who do not play by these rules, but I propose to
ignore them in what follows. Paul Smolensky, to his credit, is not among
them.
Smolensky has recently been spending a lot of his time trying to show
that, vivid first impressions to the contrary notwithstanding, some sort of
connectionist cognitive architecture can indeed account for composition-
ality, productivity, systematicity, and the like. It turns out to be rather a
long story how this is supposed to work; 185 pages of a recent collection
of papers on connectionism (Macdonald and Macdonald, 1995) are
devoted to Smolensky' s telling of it, and there appears still to be no end
in sight. It seems it takes a lot of squeezing to get this stone to bleed.
Still, Smolensky' s account of compositional phenomena has had a
good press in some quarters; the Churchlands tell us, for example, that
"Smolensky (1990) has shown that [his sort of cognitive architecture is]
at least adequate to embody . . . the systematic linguistic structures and
transformations deemed essential by Fodorean accounts of cognition"
114 Chapter 10
(McCauley, 1995, 234). It would certainly be big news if that were true.
But, in fact, Smolensky' s account doesn't work, and the formal details,
though they are daunting and take up a lot of pages in Smolensky' s
papers, are largely inessential to understanding what's gone wrong.
To begin, I want to sketch an area of consensus. Unlike the kind of
cognitive theory that Smolensky calls '1ocal" connectionism, and quite
like what have come to be called "classical" or '1anguage of thought"
2
cognitive architectures, Smolensky architectures
3
invoke a structural
relation between concepts (mutatis mutandis, between mental representa-
tions)4 to explain compositional phenomena. The accounts of composi-
tionality that S-architectures and classical theories propose are thus of the
same general form. Both tum essentially on postulating an (asymmetric)
structural relation (call it R) that holds between concepts and constrains
their possession conditions and their semantics. In particular, according to
both S-architectures and classical architectures:
(i) (possession conditions) If a concept C bears R to a concept c,
then having C* requires having C; and
(ii) (semantics) If a concept C bears R to a concept c, then the
content of C* is determined (at least inter alia) by the content of C.
We'll see presently that what distinguishes S-architectures from classi-
cal architectures is what relation they say R is.
I can now tip my polemical hand: I'm going to argue that the account
of R that S-theories give is entirely and haplessly parasitic on the account
of R that classical architectures give. That's the sense in which, even after
all those pages, Smolensky hasn't so much as made a start on constructing
an alternative to the classical account of compositional phenomena. I pro-
pose to take this in three stages: First, I want to remind you of what clas-
sical theories say about relation R. I'll then say how Smolensky' s putative
alternative is supposed to work. I'll then say why Smolensky' s alternative
is merely putative. Then I'll stop.
Stage 1: Classical Theories
According to classical architectures, R is the constituency relation: C bears
R to c iff C is a constituent of C*. It's worth getting clear on how the
assumption that R is the constituency relation connects with assumptions
(i) and (ii) and with the classical account of compositional phenomena.
The Connection with (i)
Roughly, the constituency relation is a part/whole relation: If C is a con-
stituent of C*, then a token of C is a part of every token of C*. More pre-
Connectionism and Systematicity (Continued) 115
cisely, constituency is a co-tokening relation; that is, if C is a constituent of
c, then it is metaphysically necessary that for every tokening of C* there
is a corresponding tokening of C. (Part/whole isn't, of course, the only
relation of co-tokening. For present purposes, it doesn't matter which of
these you choose as R. For discussion, see chapter 9.) Since on this con-
strual of R, nobody can token C* without tokening C, it follows that
nobody can have C* without having C.
The Connection with (ii)
Reading R as the constituency relation doesn't literally entail (ii) since
presumably it's possible for a system of representations to exhibit syntac-
tic constituency without exhibiting semantic compositionality (in effect,
all the syntactically complex symbols in such a system would be idioms).
5
But taking R to be the constituency relation does provide for a natural
explication of the informal idea that mental representations are composi-
tional: what makes them compositional is that the content of structurally
complex mental symbols is inherited from the contents of their less struc-
turally complex parts.
To claim that this is a "natural" way to explicate compositionality is
to put the case very mildly. If R is constituency, then (ii) says that the
semantics of complex representations derives from the semantics of their
parts. The effect is to make the connection between compositionality, sys-
tematicity, productivity, and the like immediately clear; an example may
serve to give the flavor of the thing.
There seems to be a strong pretheoretic intuition that to think the con-
tent brown cow is ipso facto to think the content brown. This amounts to
considerably more than a truism; for example, it isn't explained just by the
necessity of the inference from brown cow to brown. (Notice that the infer-
ence from two to prime is necessary too, but it is not intutively plausible
that you can't think two without thinking prime.) Moreover, the intuition
that thinking brown cow somehow requires thinking brown is pretty
clearly connected with corresponding intuitions about the systematicity
of the concepts involved. It's plausibly because thinking the content brown
cow somehow invokes the concept BROWN that anybody who is able to
think brown cow and tree is ipso facto able to think brown tree.
Assuming that (ii) holds and that R is the constituency relation (or
some other co-tokening relation; see above) is a way to guarantee that
thinking brown cow does indeed invoke the concept BROWN, so it pro-
vides an elegant solution to the galaxy of observations I just retailed. If
the mental representation BROWN is a part of the mental representation
BROWN COW, then of course you can't have a token of the former in
your head without having a token of the latter in your head too. System-
aticity likewise falls into place: If the mental representation BROWN is
116 Chapter 10
a part both of the mental representation BROWN COW and of the
mental representation BROWN TREE, then if you're in a position to think
the intentional (semantic) content that BROWN contributes to BROWN
COW, you are thereby in a position to think the intentional (semantic)
content that BROWN contributes to BROWN TREE . Constituency and
compositionality, taken together, guarantee that all of this is so.
Very satisfactory, I would have thought; so I want to emphasize that
none of this follows from (ii) alone (i.e., from (ii) without the identification
of R with constituency). (ii) says only that if C bears R to C*, then C (par-
tially) determines the intentional and semantic content of c. But that
doesn't come close to entailing that you can't think c without thinking
C; and, as we've been seeing, it's this latter constraint that the explanation
of compositional phenomena appears to tum on. What accounts for com-
positionality, according to classical architectures, is (ii) together with the
assumption that if C bears R to C*, then C is co-tokened with C*; and this
assumption is guaranteed, in tum, by the identification of R with con-
stituency. (More on all this later in the discussion.)
Stage 2: Smolenksy Architectures
For present purposes, the briefest sketch will do.
According to Smolensky, mental representations are vectors: "[T]he
import of distributed representation is precisely that ... a representation is
not to be found at a unit but in an activity pattern. Mathematically, an
activity pattern is ... a list of numbers ... giving the activity of all the
units over which the pattern resides" (232). All you need to know about
such representations, f ~ r present purposes, is that they do not decompose
into constituents. Of course lists can have sublists, and every sublist of a
vector specifies the activation level of some unit(s) or other. But there is,
for example, no guarantee that the list of numbers that expresses the con-
tent brown cow and the list of numbers that expresses the content brown
tree will have any of their sublists in common. (Quite generally, the fact
that a certain vector has a semantic interpretation does not even guaran-
tee that its subvectors have semantic interpretations too. This is because
it's not guaranteed-in fact, it's not usually the case-that every "unit"
has a semantic interpretation.) We're about to see that S-architectures do
offer a relation R that makes (ii) true, but it isn't a part/whole relation; in
fact, it isn't a co-tokening relation of any kind.
So, if R isn't constituency, what is it? Roughly, according to S-
architectures, R is a derivation relation. We arrive at the core of
Smolensky's theory. The core of Smolensky's theory is an algorithm for
encoding constituent structure trees as vectors. Once again, the details
don't matter (for a sketch, see Smolensky [1995] circa 236), but for the
Connectionism and Systematicity (Continued) 117
sake of the argument, I'll suppose (with Smolensky) that the encoding
algorithm is bi-unique: Given a tree, the algorithm yields a corresponding
vector; given a vector derived from a tree, the algorithm yields the tree
from which the vector is derived.
Here's how things stand so far: Vectors don't have constituents. But
vectors can be derived from trees; and, of course, trees do have con-
stituents. Smolensky' s proposal (though he doesn't put it quite this way)
is that we introduce a notion of derived constituency for vectors that works
as follows: C is a derived constituent of vector V iff V (uniquely) encodes C*
and Cis a constituent of c. That is, the derived constituents of a vector V
are the constituents tout court of the tree that V encodes. So, for example,
the vector for brown needn't (and generally won't) be a part of the vector
for brown cow. But if (by assumption) V encodes the tree ((brown) (cow)),
then the subtree (brown) is a classical constituent of the tree that V
encodes. So it follows from the definition that the subtree (brown) is a
derived constituent of V. So, according to S-architectures, R is a sort of
constituency relation after all, only it's, as it were, a constituency relation
once removed: Though it's not a part/whole relation over vectors, it is a
part/whole relation over what the vectors encode.
As you will probably have gathered, I don't actually care much whether
this apparatus can be made to operate (e.g., whether there is an algorithm
of the sort that S-architecture wants to use to encode vectors into trees
and vice versa). But I am very interested in the question of whether, if the
apparatus can be made to operate, then a connectionist architecture can
explain the compositionality. So let's tum to that.
Stage 3: Why Smolensky's Solution Still Doesn't Work
To begin, I want to reiterate a point that chapter 9 insisted on. The
explanations of compositional phenomena that classical theories offer are,
in general, quite naturally construed as causal explanations. That this is so
follows from the classical construal of R as constituency. Constituency is,
as previously remarked, a co-tokening relation; so if BROWN is a con-
stituent of BROWN COW, it can hardly be surprising that whenever a
token of the latter is available to play a causal role in a mental process, so
too is a token of the former. This is the heart of the classical account of
compositionality: If the effects of tokening BROWN COW partially
overlap with the effects of tokening BROWN TREE, that's because, in a
perfectly literal sense, the causes of these effects do too.
Now, patently, this pattern of explanation is not availabl to an S-
architecture; remember, the tokening of the BROWN COW v ~ c t o r does
not guarantee a co-tokening of the BROWN vector, so it can't be taken
for granted that the effects of a tokening of BROWN COW will include
118 Chapter 10
the effects of a tokening of BROWN. Mental representations are, by
definition, necessarily co-tokenend with their classical constituents. But it
is not the case that they are, by definition or otherwise, necessarily co-
tokened with their derived constituents. In fact, it's perfectly possible that
a vector should be tokened even if none of its derived constituents ever was
or ever will be. Remember that all that connects the vector V with its
derived constituents is that there is a biunique algorithm which, if it were
executed, would compute a tree such that the derived constituents of the
vector are the real constituents of the tree. But nothing in the notion of
an S-architecture, or in the kind of explanations S-architectures provide,
requires that such subjunctives ever get cashed. It is perfectly alright for C
to be a derived constituent of vector token V, even though all the tokens
of C are, and always will be, entirely counterfactual.
Now, I suppose it goes without saying that merely counter/actual causes
don't have any actual effects. So explanations that invoke the derived
constituents of a vector cannot, in general, be construed as causal expla-
nations. So, as chapter 9 wanted to know, if the explanation of composi-
tonality that S-architectures offer isn't causal, what sort of explanation is it?
Smolensky has said a lot about this in his papers, but I honestly don't
understand a word of it. For example: ''Tensor product constituents [what
I'm calling "derived" constituents] play absolutely indispensable roles in
the description and explanation of cognitive behavior in ICS [integrated
connectionist symbolic architectures]. But these constituents do not have
a causal [sic] role in the sense of being the objects of operations in algo-
rithms actually at work in the system. The5e constituents are in this sense
acausally explanatory [sic]
11
(249). If, like me, you don't know what sort
of thing the acausally explanatory explanation of "cognitive behavior"
would be, I don't expect you'll find that such passages help a lot, since
though they tell you what acausal explanations aren't (viz., they aren't
causal), they don't tell you what acausal explanations are. There are, notice,
quite a lot-say, infinitely many-of domains of "objects" that aren't
ever tokened in anybody' s head but which correspond hi-uniquely to the
representations that mental processes work on; every scheme for Gooel
numbering the mental representations provides one, as Smolensky himself
observes. Do all these "objects" acausally explain cognitive behavior? If
not, why not?
Digression on singing and sailing In an enigmatic footnote to the passage I
just quoted,
6
Smolensky asks rhetorically whether it is "so incomprehen-
sible that Simon's and Garfunkel's voices each have causal consequences,
despite the fact that neither are 'there' [sic] when you look naively [sic] at
the pressure wave realizing The Sounds of Silence?" (284)
7
What makes this
puzzling is that the passage in the text wanted to deny what the footnote
Connectionism and Systematicity (Continued) 119
apparently wishes to assert; namely, that theories that appeal to derived
constituents thereby invoke causes of the events that they explain. Smo-
lensky himself appears to be having trouble figuring out which explana-
tions are supposed to be the "acausal" ones. I do sympathize.
In any event, the answer to the rhetorical question about Simon and
Garfunkel is "yes and no." That the several voices should have their
several effects is not incomprehensible so long as you assume that each of the
voices makes its distinct, causal contribution to detenpining the character of the
waveform to which the ear responds. Alas, the corresponding assumption is
not true of the contribution of trees and their constituents to the values of
the vectors that encode them in S-architectures.
Perhaps a still homelier example will help to make things clear. One
normally sails from A to B along a vector that corresponds (roughly)
to the sum of the force that the wind exerts on the boat and the force
that the current exerts on the boat. The wind and the current make their
respective causal contributions to determining your track jointly and
simultaneously in the ways the resultant vector expresses. The counter-
factuals fall into place accordingly: Increase the force of the wind and,
ceteris paribus, the value of the vector changes in the wind's direction;
increase the force of the current and, ceteris paribus, the value of the vec-
tor changes in the current's direction. And likewise, mutatis mutandis, for
Simon's and Garfunkel' s respective contributions to determining the form
of the sound wave your hi-fi set produces and your ear responds to. This
is all fine, well understood, and not at all tendentious.
But notice that this story is quite disanalogous to the way in which, on
Smolensky' s account, vector values vary with the constituent structures
of the trees that they encode. For, to repeat, the wind and current deter-
mine the value of the vector that one sails along by exerting their respective
causal forces on the boat. So, it goes without saying, there must really be
winds and currents acting causally upon the hull, and there must be causal
laws that control these interactions. Otherwise, that the boat sails along
the vector that it does is an utter mystery. It's only because properties like
being a force four wind are occasionally instantiated and locally causally
active that the vectors one sails along ever have the values that they do.
If the wind and the current were just "imaginary"-specifically, causally
nugatory-boats without motors wouldn't move.
In contrast, remember, the trees that vectors encode need never be
tokened, according to Smolensky's theory. A fortoiri, the theory does not
countenance causal interactions that involve tree tokens, or causal laws
to govern such interactions. The question thus arises for Smolensky (but
not for Simon or Garfunkel or for me out sailing) is how-by what causal
mechanism-the values of the putative vectors in the brain come to
accord with the values that the encoding algorithim computes. On the one
120 Chapter 10
hand, Smolensky agrees that explaining compositionality (etc.) requires
tree-talk. But, on the other hand, he concedes that S-achitectures acknowl-
edge no causal transactions involving trees. Smolensky really does need
an explanation of how both of these claims could be true at the same time.
But he hasn't got cne.
To return to the main thread: Let's grant, to keep the argument going,
that there are indeed acausal explanations, and that derived constituents
are sometimes invoked in giving them. Granting all that, can't we now
stop the philosophy and get to the cognitive science? No, Smolensky is
still not out of the woods. The problem resides in a truism that he appears
to have overlooked: Bi-uniqueness goes both ways.
Apparently Smolensky is reasoning as follows: That vectors are hi-
uniquely derivable from trees licenses an architecture according to which
vectors causally explain some of the facts and trees acausally explain the
rest. But then, by the very same reasoning, that trees are hi-uniquely
derivable from vectors should license an architecture according to which
trees causally explain some of the facts (compositional phenomena, as it
might be) and vectors acausally explain the rest. Sauce for the goose,
sauce for the gander, after all. If unique derivability in one direction suf-
fices to transfer the explanatory power of a tree architecture to vectors,
unique derivability in the other direction should correspondingly suffice
to transfer the explanatory power of a vector architecture to trees; prima
facie, the total of explanation achieved is the same in either direction.
9
I
am not, please note, offering this observation either as an argument for
tree architectures or as an argument against vector architectures; it's just a
reductio of Smolensky's assumption-which is, anyhow, wildly implau-
sible on the face of it-that explanatory power is inherited under unique
derivability.
The plot so far: Smolensky thinks the best account of compositionality
is that it's causally explained by vectors and acausally explained by trees.
But he explicitly supposes that trees and vectors are biuniquely derivable,
and apparently he implicitly supposes that explanatory power is inherited
under biunique derivation. So it looks like, on Smolensky's own princi-
ples, the account according to which compositionality is causally explained
by trees and acausally explained by vectors must be as good as the one
that Smolensky actually endorses. Smolensky needs, to speak crudely, to
get some asymmetry into the situation somewhere. But there doesn't
seem to be any place for him to put it.
Patently, Smolensky can't afford to give up the inheritance of explana-
tion under biunique derivability. That would invite the question of why
the derivability of vectors from trees should be any reason at all for sup-
posing that vectors can explain what trees do. Likewise, he can't afford to
Connectionism and Systematicity (Continued) 121
give up the biuniqueness of the algorithm that encodes trees as vectors.
For, on the one hand, if you can't derive a unique vector corresponding to
each tree, then clearly the vector notation doesn't preserve the structural
distinctions that the tree notation expresses; this is trivially true, since if
trees map onto vectors many to one, the vectors thereby fail to preserve
the structural properties that distinguish the trees. And, on the other
hand, if you can't derive from each vector the unique tree that it encodes,
then the concept of a "derived constituent" of a vector isn't well defined
and the whole scheme collapses.
But even though Smolensky needs explanation to be inherited under
derivability and "is derivable from," to be symmetric, "is (actually) derived
from" needn't be. So why, you might wonder, doesn't Smolensky say this:
''The reason vectors explain what trees do but trees don't explain what
vectors do is that vectors are actually derived from trees in the course of
mental processing, but not vice versa." This would have the added virtue
of making "derived constituent of" a sort of co-tokening relation: to get
a vector token for mental processes to operate upon, you would need
first to token the tree that the vector derives from. And if derived con-
stituency is a co-tokening relation after all, then maybe S-architectures
could provide for causal explanations of compositionality phenomena
after ali-in which case, Smolensky could just scrap the puzzling talk
about "acausal explanations"-a consummation devoutely to be wished.
In fact, however, Smolensky doesn't and mustn't say any of that. For
the mind actually to derive vectors from trees....,....for it actually to execute
the algorithm that takes trees onto vectors-would require a cogni-
tive architecture that can support operations on trees. But, by general
consensus, S-architectures can't do so; S-architectures work (only) on
vectors. That, after all, is where we started. It was, as we saw, because
S-architectures don't support operations on trees that they have trouble
explaining compositionality in the first place.
To put the same point slightly differently: If Smolensky were to sup-
pose that vector tokens come from tree tokens via some algorithm that
the mind actually executes in the course of cognitive processing, he
would then have to face the nasty question of where the tree tokens
themselves come from. Answering that question would require him to
give up his connectionism since, again by general consesus, connectionist
architectures don't generate trees; what they generate instead is vector
encodings of trees.
It's important, at this point in the discussion, to keep clear on what role
the tree encoding algorithm actually plays in an S-architecture.
10
Strictly
speaking, it isn't part of Smolensky' s theory about the mind at all; strictly
speaking, it's a part of his theory about theories about the mind. In particu-
lar, it's a device for exhibiting the comparability of classical architectures
122 Chapter 10
and S-architectures by translating from the tree vocabulary of the former
into the vector vocabulary of the latter and vice versa. But there is no
cognitive architecture that postulates mental processes that operate on
both kinds of representation (at least there is none that is party to the
present dispute);
11
a fortiori, there is no mental process in which the two
kinds of representations are supposed to interact; and no mind ever
executes Smolensky' s encoding algorithm in the course of its quotidian
operations (except, perhaps, the mind of a theorist who is professionally
employed in comparing classical architectures with connectionist ones).
So if everybody, Smolensky included, agrees that the encoding algo-
rithm isn't really part of Smolensky's theory of how the mind works, why
does Smolensky keep making such a fuss about it? It's because, since he
admits that the explanation of compositionality should be couched in
terms of trees and their constituents, Smolensky needs somehow to make
the vocabulary of tree-talk accessible to vector theories. The function of
the encoding algorithm in Smolensky' s overall picture is to permit him to
do so, and hence to allow the connectionist explanation of composition-
ality to parasitize the classical explanation. That's all it does; it has no
other motivation.
The sum and substance is this: Smolensky' s argument is, for all intents
and purposes, that since there is exactly one vector that is derivable from
each tree, then if the structure of a tree explains compositionality (or
whatever else; the issue is completely general), so too does the structure
of the corresponding vector. Smolensky gives no grounds, other than their
biunique derivability from trees, for claiming that vectors explain what trees do.
Put this way, however, the inference looks preposterous even at first
blush; explanation is not, in general, preserved under one-to-one correspondence;
not even if the correspondence happens to be computable by algorithm.
Why on earth would anyone suppose that it would be?
In effect, Smolensky proposes that the classical theory should de the
hard work of explaining compositionality, systematicity, etc., and then the
connectionist theory will give the same explanation except for replacing
"constituent" with "derived constituent" and "explain" with "acausally
explain" throughout. Would you like to know why thinking brown cow
requires thinking BROWN? Glad to oblige: It's because, since BROWN is
a classical constituent of BROWN COW, it follows by definition that the
BROWN vector is a derived constituent of the BROWN COW vector.
And, by stipulation, if c is a derived constituent of c 1 then your thinking
c acausally explains your thinking C. Smolensky' s architecture offers no
alternative to the classical story and adds nothing to it except the defi-
nition and the stipulation. This way of proceeding has, in Russell's famous
phrase, all the virtues of theft over honest toil. Can't something be done
about it?
Connectionism and Systematicity (Continued) 123
What Smolensky really wants is to avail himself of tree explanations
without having to acknowledge that there are trees; in effect, to co-opt
the vocabulary of the classical theory but without endorsing its ontology.
But that, alas, he is not permitted to do, unless he is prepared to recognize
the resulting theory as merely heuristic. (It is, of course, perfectly alright
to talk about the Sun's orbit around the Earth for purposes of doing navi-
gation; but that's sailing, not science.)
If, for an example that I hope is untendentious, you want rock-talk in
your geological theory-if you want, that is, to frame your geological
generalizations and explanations in terms of rocks and their doings-you
will have to admit to rocks as part of the actual causal structure of the
world. What you are not allowed to do is borrow rock-talk when you
want to explain what it is that holds Manhattan up and also endorse a rock-
free ontology when you come to saying what the world is made of; that would
be cheating. After all, how could rocks explain what holds Manhattan up if
there aren't any rocks? It is not, notice, a respectable way of avoiding this
question to reply that they do hold it up, only acausally.
Likewise, you are not allowed to borrow the idea that the constituent
structure of classical mental representations is what explains the composi-
tionality of thought and also deny that there are mental representations
that have classical constituent structure; that is cheating too. How could
classical constituents explain why thought is compositional if thoughts
don't have classical constituents?
Smolensky proceeds as though the choice of an explanatory vocabulary
and the choice of an ontology were orthogonal parameters at the cogni-
tive scientist's disposal. To the contrary: What kinds of things a theorist
says there are sets an upper bound on what taxonomy his explanations
and generalizations are allowed to invoke. And what taxonomy his
explanations and generalizations invoke sets a lower bound on what kinds
of things the theorist is required to say that there are. In this fashion,
science goes back and forth between how it claims the world works and
what it claims the world is made of, each kind of commitment constrain-
ing the other (to uniqueness if we're lucky).
Smolensky, it appears, would like a special dispensation for connec-
tionist cognitive science to get the goodness out of classical constituents
without actually admitting that there are any. In effect, he wants, just this
once, to opt out of the duality of ontology and explanation; that's what
his appeals to acausal explanation are intended to allow him to do. It's the
special convenience of acausal explanations that, by definition, they carry
no ontological burden; just as it's the special convenience of free lunches
that, by definition, there is no charge for them. That's the good news. The
bad news is that there aren't any.
124 Chapter 10
Acknowledgment
I'm very grateful to Zenon Pylyshyn for his comments on an earlier draft
of this chapter.
Notes
1. Just to have a label, I'D sometimes call whatever things belong to this bundle "composi-
tional phenomena." It won't matter to my line of argument exactly which phenomena
these are, or whether compositionality is indeed at the center of the cluster. Though I
have strong views about both of these questions, for present purposes I don't need to
prejudge them.
2. For this terminology, see Fodor and Pylyshyn (1987). What distinguishes classical archi-
tecture &om local connectionism is that the former recognizes two sorts of primitive
relations between mental representations, one causal and one structural. By contrast,
local connectionist architectures recognize the former but not the latter. Fodor and Pyly-
shyn argue, correctly, that the problems that local connectionists have with composi-
tional phenomena trace to this fad. Apparently Smolensky agrees with them.
3. More terminology: Smolensky calls his kind of theory an "integrated connectionist/
symbolic cognitive architecture." That, however, is a mouthful that I'd rather not swal-
low. I'll call it a "Smolensky Architecture" or "Smolensky Theory" ("S-architedure" or
"S-Theory" for short).
4. Still more terminology. (Sorry.) For present purposes, I'll use "concept" and "mental rep-
resentation" pretty much interchangably; I'm supposing, in effect, that concepts are
interpreted mental representations. I shall, however, want to distinguish between a con-
cept (or mental representation) and its semantic value (e.g., the individual that it denotes
or the property that it expresses). I'll write names for concepts in capitals (thus "RED"
denotes the concept RED) and I'll write names for the semantic values of concepts in
italics (thus, the concept RED expresses the property red.) These conventions tacitly
assume a representational theory of mind; but that's common ground in the present dis-
cussion anyway.
The reader should bear in mind that ''RED," ''BROWN COW," and the like are sup-
posed to be names of concepts, not structural descriptions. The notation thus leaves open
whether "BROWN COW" (or, for that matter, ''RED") names a complex concept or an
atomic one. It is also left open that mental representations are kinds of vectors and have
no constituents.
5. This is concessive. If you think that, in the long run, even what passes as syntactic con-
stituency must be semantically d e f i n e d , ~ much the better for the line of argument I'm
about to develop.
6. The next three paragraphs differ substantially &om the original published version.
Smolensky has often complained that the role vectors play in his theory is just like the
role that they play in untendentious scientific explanations in (e.g.) mechanics. So why is
everybody picking on Smolensky7 This question deserves an answer; the discussion in
the text provides it.
By the way, the discussion here connects with what's said in passages in chapter 9:
that the "normal decomposition" of a vector is, ipso facto, decomposition into factors
the ontological and causal status of which the theory acknowledges. The sailor factors
the vector he sails on into components for wind and tide because he believes (correctly)
that both are causally implicated in determining his course.
I am indebted to a discussion with Smolensky, Brian Mclaughlin, and Bruce T essor
among others for these revisions.
Connectionism and Systematicity (Continued) 125
7. I have made inquiries. It would appear that Tht Sounds of Silence is some popular song or
other, of which Simon and Garfunkel are among the well-known expositors. I don't
imagine that the details matter much.
8. Sometimes Smolensky writes as though S-architectures offer reductions of classical archi-
tectures (see, e.g., 1955b, 272: "the symbolic structure of [classical] representations and
the recursive character of the functions computed over these representations have been
reduced to [my emphasis] tensor product structural properties of activation vectors .... "
See also 1995a, passim). But that can't really be what he intends. Causality is preserved
under reduction; it couldn't be that water reduces to H
2
0 and that H
2
0 puts out fires
and that water doesn't. But that vectors are causes and trees aren't is Smolensky's main
claim.
Chapter 9, trying hard to square what Smolensky says about reduction and expla-
natory levels with his insistence that classical explanations are acausal, suggests that
maybe Smolensky is really some sort of closet epiphenomenalist. I still wouldn't be in
the least surprised.
9. To be sure, the two treatments would differ in the way they divide the burden between
causal and acausal explanation; and in principle, that might allow one to choose between
them. Perhaps, for example, some kinds of fads are intrinsically suited for causal expla-
nations, while others are, by their nature, best acausally explained. But who knows
which kinds of fads are which? In particular, what could justify Smolensky in assuming
that fads about compositionality are of the latter sort7
10. I think, by the way, that Smolensky is clear on this; that is why he conceeds that "in the
classical, but not the ICS architectures ... constituents have a causal role in processing"
(236). I therefore regard this part of the discussion as untendentious and expository. The
polemics start again in the next paragraph.
11. Why not a "mixed" theory, as a matter of fad7 Sure, why not. But the claim that some
mental processes are vector transformations sounds like a lot less than the pardigm shift
that coMectionists keep announcing. I suppose, for example, Smolensky would agree
that if a cognitive architecture allows tree operations as well as vector operations, it
should be the former that its explanation of compositional phenomena appeals to. It is,
after all, exactly because he admits that trees are the natural way to explain composi-
tionality that Smolensky feels compelled to invoke explanations in which they figure
acausally.
Chapter 11
There and Back Again: A Review of Annette
Karmiloff-Smith' s Beyond Modularity
These days, hordes of people are interested in the idea that aspects of
cognitive architecture may be modular. I know at least two or three
(people, not hordes or aspects), and there may be others. But "modu-
larity" means different things on different tongues. In this chapter, I want
brieRy to distinguish between some versions of modularity theory that
are currently in play. Then, I will discuss one of them in critical detail.
There are four essential properties connected with the notion of a
module: Unless you believe that at least some mental entities instantiate
at least two of them, you are not in the modularity camp according to my
way of choosing sides.
1. Encapsulation Information flow between modules-and between
modules and whatever unmodularized systems the mind may contain-is
constrained by mental architecture. "Constrained by mental architecture"
means "not cognitively penetrable" (Pylyshyn, 1984): You can't change
such an arrangement (just) by fooling around with someone"s beliefs and
desires. In particular, architectural arrangements are (relatively) insensitive
to instructional variables in experimental tasks. The persistence of illu-
sions is the classical instance. Convincing the subject that the Muller-Lyre
effect is illusory doesn't make the apparent difference between the length
of the lines go away.
2. Inaccessibility In effect, the inverse of encapsulation. Just as informa-
tion about beliefs and desires can't get into a module, so the information
that is available to its computations is supposed to be proprietary and
unable to get out. In particular, it is supposed not to be available for the
subject's voluntary report.
3. Domain specificity The information and operations by which a module
is constituted apply only in the module's proprietary domain. Concepts
and processes may thus be available for language learning, or face recog-
nition, or space perception which are not likewise available for balancing
one's checkbook or deciding which omnibus to take to Clapham.
128 Chapter 11
Table 11.1
NC AKS
JAF
encapsulated don't care
yes and no# yes
inaccessible yes yes and no% yes
domain-specific yes yes yes
iMate yes yes and no@ yes
Note: NC is Chomsky, AKS is Kanniloff-Smith, and JAF is me.
4. Innateness The information and operations proprietary to a module
are more or less exhaustively "genetically preprogrammed" (whatever,
exactly, that means).
People who agree that some mental processes are modular may, none-
theless, differ appreciably in their views about the encapsulation, accessi-
bility, domain specificity, and innateness of even their favorite candidates.
Table 11.1 shows a rough sketch of the way three currently active theo-
rists distribute, all of whom think of themselves as promodule in some
sense or other. Comments:
"'Chomsky, in some of his moods, dislikes the whole information-
processing view of mental operations. If the mind isn't an information
processor at all, then the question of whether it's an encapsulated infor-
mation processor doesn't arise.
#A proposed principle of the ontogeny of cognition: Mental processes
become encapsulated in the course of cognitive development (perhaps
through overlearning). So they are encapsulated synchronically but not
diachronically. I'll refer to this putative process as modulariZiltion (my term,
not Karmiloff-Smith's).
%Another proposed principle of cognitive development: Modularized
information becomes increasingly accessible over time as a result of an
"epigenetic" process of representational redescription (the "RR" theory).
1
I'll
refer to this as a process of demodularization (again, my term).
@What's innate: Some domain-specific information and "attentional
biases"; and, presumably, the psychological mechanisms that underlie the
putative epigenetic processes. But neither encapsulation nor accessibility
are themselves genetically preprogrammed.
An aside about attention: It's a recurrent theme in Beyond Modularity
(BM) and also in Elman et al. (see chapter 12) that ''There must be some
innate component to the acquisition of [e.g.] language [but] ... this does
not mean that there has to be a ready-made module. Attention biases and
some innate predispositions could lead the child to focus on linguistically
relevant input and, with time, to build up linguistic representations that
are domain specific" (36). This emphasis on innate attentional biases is not
There and Back Again 129
widely shared by modularity theorists. It strikes me as unpromising, and I
won't discuss it in ' what follows. In neither Karmiloff-Smith' s book nor
Elman's is it explained how one could have a disposition (innate or other-
wise) to concentrate on Xs unless one already has the concept of an X.
(''Pray, attend to the passing flubjumbs." "Can't." 'Why notr' "Don't
know what a flubjumb is.") Postulating innate attentional biases doesn't
dispense with the postulation of innate conceptual content; it just presup-
poses it.
It may be that, in passages like the one just quoted, Karmiloff-Smith is
only suggesting that it would be a help to the child to be (differentially)
interested in speechlike sounds. That's quite plausible, in fact; but it doesn't
even begin to explain how someone who is so biased manages "with time,
to build up linguistic representations that are domain speci6c." As far as
anybody knows, you need innate conceptual content to do that; indeed,
as far as anybody knows, you need great gobs of it. (I am disposed to
attend to the speech sounds that German speakers make; but I And learn-
ing German very hard for all that.)
So much for some current kinds of modularity theories. Perhaps I
should say at the outset that I think you'd have to be crazy to bet much
on which, if any, of them is true. The study of mental architecture is in its
infancy, and it looks to be developing very slowly. My modest ambition in
what follows is just to indicate some doubts about Karmiloff-Smith' s
view. And, even here, I'm not going to argue for anything so positive or
decisive as that she is plain wrong about modularity. I will, however, try
to show that the ways she sets out her view, and the ways she undertakes
to defend it, are insensitive to certain distinctions that a cognitive archi-
tect really ought to observe. And that, when this is all cleared up, what's
left may after all be true--who knows?-but, as things stand, neither
the arguments for modularization (the thesis that cognitive architecture
becomes increasingly modular with development) nor the arguments for
demodularization (the thesis that information in modules becomes increas-
ingly accessible with development), are persuasive.
Modularization
I won't treat modularization at length since Karmiloff-Smith makes practi-
cally no positive case for it except for remarking that the plasticity of the
infant's brain militates against the thesis that its cognitive architecture is
innately preformed. I think this consideration cuts little ice against nativ-
ism, since modularization, if there is such a process, might be matura-
tional, and the course of maturation might itself be genetically speci6ed.
(Something like this is surely true for the development of secondary
sexual characteristics, for example; why, then, couldn't it be true of brain
130 Chapter 11
structures?) And anyhow, nobody knows what the neural plasticity of
the infant's brain means. Nobody has any idea, for example, whether the
infant's brain is plastic in respects that affect cognitive architecture. (For more
on this, see chapter 11.)
No doubt, Karmiloff-Smith is right to insist that nothing we know
actually rules out progressive modularization as a process in cognitive
development. As she says, though "[i]t is true that some [sic] genetically
specified predispositions are likely to be involved ... such a claim should
not automatically negate the epigenetic [sic] influence of the sociocultural
environment on the [child's cognitive] development" (129). As far as I can
see, however, there is no positive evidence that a processess of modulari-
zation does, in fact, occur in ontogeny; and, lacking such evidence, steady-
state is surely the least hypothesis.
Karmiloff-Smith does, however, make a case for demodularization;
doing so occupies most of her book. I now turn to this.
Demodularization
It's essential to Karmiloff-Smith' s story that there are interesting, endoge-
nously driven reorganizations of cognitive domains that typically occur
after a child achieves "behavioral mastery" in the domain-hence, after
the point at which most developmental cognitive psychologists lose
interest in cognitive development. Among these is a purported increase in
the accessibility to voluntary report of information that was previously
propreitary to a modularized system. Here's one of her parade examples:
On the one hand, "once young children are beyond the very initial stage
of language acquisition and are consistently producing both open-class
and dose-class words . . . there can be no question that at some level
these are represented internally as words [sic]" (51). But, on the other
hand, 'When asked to count words in a sentence, young children fre-
quently neglect to count the closed-class items. When asked directly
if 'table' is a word, they agree; but when asked if 'the' is a word, they
answer in the negative" {51-52). What explains such findings, according
to Karmiloff-Smith7 ''The RR model posits this developmental progres-
sion can be explained only by invoking, not one representation of lin-
guisic knowledge, to which one either has or does not have access, but
several re-representations of the same knowledge, allowing for increasing
accessibility" (54).
This sort of case is very close to the heart of Karmiloff-Smith' s attempt
to get beyond modularity, so I want to look at it rather closely. I'll make
two claims, which will form the substance of the discussion to follow:
first, there is actually no evidence that the accessibility of modularized
information increases over the course of cognitive development; and
There and Back Again 131
second, even if accessibility does increase, the redescription of the modu-
larized information wouldn't explain why it does.
Let's start with the claim that the six-year-old child, who explicitly
reports that "the" is a word, has access to intramodular representations
that are inaccessible to the three year old. The three year old, we're
assuming, marks the word boundary between "the" and ''boy" in the
course of such modularized "on-line" tasks as parsing an utterance of "the
boy runs," but nonetheless denies that "the" is a word when asked for
metalingustic judgments. So, plausibly, something is accessible to the older
child that the younger one can't get at. The crucial question is whether it's
the accessibility of intramodular representations-that is, of information
inside the module-that has changed in the course of development.
(In passing: I say it's plausible that it's something about the accessibility
of linguistic information that changes between ages three and six, not that
it's apodictic. It could be that all that happens is that a certain linguistic
confusion gets resolved: Young children think the word "word" means
open-class word, whereas older children know better. It wouldn't be
awfully surprising if three year olds are confused about this; most words
are open-class words, after all, and it's likely that the examples that the
child learns the word "word" from are themselves open-class. "Cat" and
"mat," unlike "and" or "of," are prototypical words, and there's plenty of
evidence that children generallly learn to identify relatively prototypical
instances of a category before they learn to identify relatively marginal
examples. I propose, however, to be concessive for the sake of the argu-
ment, and just take for granted that the developmental change has some-
thing to do with alterations of accessibility.) .
There is, nonetheless, a perfectly plausible alternative to the theory that
what has become more accessible is an intramodular representation. To
see this, let's consider for a moment what the putative language module
does when it's running as an input parser. I assume-and I assume that
Karmiloff-Smith would do so too-something like the picture in figure
11.1. The function of the language parser is to map incident utterance
tokens, specified acoustically, onto the sorts of objects that linguists call
"structural descriptions." just what a structural description is depends, to
a certain extent, on which linguist you talk to, on which day of the
input sentence - PARSER -so
Figure 11.1
132 Chapter 11
week. But there's a broad consensus that it must consist in a mental rep-
resentation of the input utterance at several levels of linguistic description.
Standard candidates for the relevant levels include the following bare
minimum: a phonetic representation, a phonological representation, a
lexical representation, and a representation of a syntactic "constituent
structure" tree.
As I say, this list is a bare minimum. But it will do for our present pur-
poses, given just one further architectural assumption (which is also not in
dispute in the current discussion), namely, that the parser computes the
structural descriptions of utterances of sentences on the basis of its inter-
nally represented grammatical information about the language from which
the sentences are drawn. For example, the parser computes that the struc-
tural description of "the boy ran" is (something like):
(((#THE# )o (#BOY# )N)NP ( # RAN)vp )s
It is able to perform this computation because it internally represents such
information as that "the" is an article, that "boy" is a noun, that "ran" is an
intransitive verb, that sequences of the form (DN) are noun phrases, etc.
Presumably, language learning consists in large part in supplying the
module with this sort of information. (I leave open, for the moment,
whether grammatical information in the parsing module is represented
"explicitly" or "procedurely"; we'll return to this issue presently.)
So much for background. Consider now the question whether the
three-year-old child who has achieved ''behavioral mastery" (e.g., who is
able fluently to parse utterances of the sentence "the boy ran") has off-line
access to the information that "the" is a word. The point to which I want
to draw your attention is that this question is ambiguous. For, according
to the present picture, the information at issue is represented twice in the
three-year old's head. On the one hand, each time the child succeeds in
parsing a sentence that contains "the," he gets a structural description that
says that "the" is a word in that sentence. And, on the other hand, part of
the information about English, that is presumably speci6ed by the gram-
mar that's in the parser, is a lexicon (i.e., a list of the English words the
child knows) which includes the word "the." So now, when you ask
whether the child has off-line access to his internal representation of the
fact that "the" is a word in "the boy ran," do you mean the representation in
the module, or do you mean the representation in the structural description?
Or, to put the same question the other way around, when the six-year
old exhibits explicit metalinguistic awareness that "the" is a word in "the
boy ran," has the change occurred in his off-line access to what's in the
module, or in his off-line access to what's in the structural description?
This matters a lot for Karmiloff-Smith, because it's only if the right answer
is the 6rst one that she has produced a bona 6de case of developmental
There and Back Again 133
demodularization. Suppose, on the other hand, that what happens in cog-
nitive development is that over time children get increasing off-line access
to the content of the structural descriptions that their modules compute.
Then, for all the arguments we've got so far, there is no such thing as
demodularization, and hence there is no developmental alteration of
intramodular accessibility for a representational redescription theory to
explain.
Now, I don't know for sure which, if either, of these stories is true. It
seems clear, in fact, that neither could be the whole truth, because neither
explains why the shift of accessibility of information about word bound-
aries should be specific to closed-class items.
2
For what it's worth, how-
ever, here are two straws in the wind to think about.
1. Off-line access to information in structural descriptions is appar-
ently top-down even for adults. For example, S' s response to sylla-
bles is faster than it is to phones in phoneme monitor tasks (Savin
and Bever, 1970). It wouldn't be surprising if ontogeny recapitulates
this asymmetry. If so, then you'd expect children to be able to
report relatively high-level facts about sentences (e.g., facts about
their grammatical structure and meaning) before they can report rel-
atively low-level facts about sentences (e.g., facts about where the
word boundaries are).
2. Information about word boundaries is represented both in struc-
tural descriptions and in the module. That's why Karmiloff-Smith' s
observations about children's metalinguistic access to word bound-
aries provides only equivocal support for demodularization. But
there is some information that is represented in the parser but not in
the structural descriptions that it outputs: for example, specifications
of grammatical rules. Thus the structural description of "the window
was broken by the rock" tells you that it's a passive; but it doesn't
tell you by what operations this (or other) passives are constructed.
3
My point is that facts about the accessibility of such information
provide a crucial test for the demodularization thesis. If what's
going on in development is that module-internal information is
becoming generally accessible, that should be true inter alia of
module-internal information that is not represented in structural
descriptions. In fact, however, the data are otherwise. Information
that is plausibly in the module but not in structural descriptions
never becomes accessible for metalinguistic report; not at six and not
at sixty either. If you want to know how passivization works, you
have to take a linguistics course.
Notice that this argument goes through even if it's assumed that
the information that's in the module is procedural (that it's ''know
how" rather than ''know that"). English speakers can't report a
134 Chapter 11
procedure for forming passives any more than they can report a rule
for doing so.
The bottom line, so far, is that although it might be true that the walls
of modules become relatively transparent in the course of cognitive
development, there isn't actually any positive reason to believe that they
do. In all the cases I can think of-and, as far as I can discover, in all the
cases that Karmiloff-Smith discusses-the data don't distinguish the
theory that it's information in the modules that becomes available for
report from the theory that it's information in the modules' output that
does. And, when there are data that do bear on this distinction, they favor
the latter hypothesis over the former.
Redescription
Suppose, however, that I'm wrong about all this: Suppose that module-
internal information does become increasingly accessible over time. What
about the idea that it is some epigenetic process of "representational
redescription" that effects this change?
This idea is Karmiloff-Smith's leitmotif. I've already quoted her remark
about the putative shift in accessibility of word boundaries, that "the RR
model posits this developmental progression can be explained only by
invoking, not one representation of linguisic knowledge, to which one
either has or does not have access, but several re-representations of the
same knowledge, allowing for increasing accessibility." This sort of point
is reiterated throughout her book. Thus, discussing the modularity, or
otherwise, of the child's theories about the minds of conspecifics, she
remarks that "the general process of representational redescription oper-
ates on the domain-specific representations of [the] theory of mind ... just
as it does in other domains of cognition, to tum [them] into data struc-
tures available to other processes .... " (136-137). And, discussing the
child's implicit theory of the mechanics of middle-sized objects, she says,
''My point is that coherently organized information about [physical]
objects is first used by the infant to respond appropriately to external
stimuli. [But] despite its coherence, it does not have the status of a
'theory.' To have theoretical status, knowledge must be encoded in a
format usable outside normal input/output relations. It is these rede-
scriptions that can be used for building explicit theories" (78). In short, "a
crucial aspect of development is the redescription of . . . knowledge into
different levels of accessible, explict formats" (163). Summarizing all this,
she says that "the human mind exploits its representational complexity by
re-representing its implicit knowledge into explicit form" (191).
Now, I admit to finding such passages hard to construe. Here's my
problem: Karmiloff-Smith appears to be claiming that it's the child's
There and Back Again 135
changing of his representational formats-in particular his changing from
formats that are less accessible to formats that are more so-that explains
the putative ontogenetic changes in the availability of intramodular
information to explicit report, and generally, to operations outside the
module's proprietary domain. But that can't be right as it stands; differ-
ences in format, in and of themselves, can't explain differences in accessi-
bility. The point is sufficiently banal. Is the information that the cat is on
the mat more, less, or equally accessible when it's expressed in English
than when it's expressed in French? That's a nonsense question on the face
of it; it's more accessible if you're a monolingual English speaker, but it's
less accessible if you're a monolingual French speaker, and it's equally
accessible if you are Auently bilingual. No "format" is either accessible or
inaccessible as such. So no story about changing formats is, in and of
itself, an explanation of changes in accessibility.
Likewise for explicitness. Every representation is explicit about some-
thing; pictures are explict about (e.g.) shape and color, sentences are
explicit about (e.g.) tense and negativity. Correspondingly, if you change
format from pictures to words, some aspects of what's represented become
less explicit, and some aspects become more so. But no system of repre-
sentation is explict or inexplicit as such; no system of representation is
more-or less-explicit than any other across the board.
Representational formats, in short, aren't "accessible" or "explicit" per
se; they're only "accessible" to something or "explicit" about some
or other of the information that they represent. So, suppose that the
child does change representational formats in the course of cognitive
development. That wouldn't explain the (putative) fact that intramodular
information becomes increasingly available in the course of cognitive
development; it just raises the question of why the information is more
accessible in the new format than it was in the old one. Indeed, it raises
the question of why the (putative) change of the representational format
should affect the accessibility of the representations at all.
I don't think that Karmiloff-Smith ever faces these sorts of questions
squarely, but she does drop a number of hints; they're implicit (or, any-
how, inexplicit) in the several by no means coextensive claims that she
makes about what kinds of redescriptions go on in the course of develop-
ment. I'll discuss four of these. The moral will be that none of Karmiloff-
Smith' s stories about how intramodular representations are redescribed in
the course of ontogeny provides a plausible account of why such repre-
sentations should become increasingly accessible over time.
1 "Procedural" Representation
In the early stage of the development of a domain-specific cognitive system,
the information relevant to task performance is encoded "procedurally."
136 Chapter 11
Thus "at level I, representations are in the form of procedures
for analyzing and responding to stimuli in the external environment ....
Information embedded in Ievel-I representations is therefore not available
to other operators in the cognitive system .... A procedure as a whole [sic]
is available as data to other operators; however its component parts [sic] are
not" (20).
It's again not clear to me how Karmiloff-Smith wants this story to go. It
sounds as though she thinks that it's the (presumed) fact that the informa-
tion is procedurely represented that explains why its component parts are
not "available to other operators in the cognitive system." But, surely, it
doesn't.
Though it gets thrown around a lot in cognitive science, the notion of a
procedural representation isn't itself transparent. At its least tendentious,
however, a procedural representation is just a representation of a proce-
dure. This is the construal that's suggested by examples like sentence
parsing. It may be that what underlies the child's ability to assign syn-
tactic forms to utterances is something like an algorithm for mapping
sentence tokens under acoustic representation onto their structural
descriptions. A grammar of the child's language may be "explicit" in the
parser, or it may be merely "implicit" in the algorithm in the sense that,
whereas the latter contains "declarative" representations (like "'the' is a
word"), the former contains "imperative" representations (like "if you find
a phonological sequence # /t/ /h/ /e/ #, label it a token of the word-type
"the"). Notoriously, parsers and grammars needn't be trivially inter-
convertible. Going in one direction, there are grammars for which parsers
aren't constructible; and, going in the other direction, there needn't be any
fact of the matter about which of an infinity of extensionally equivalent
grammars a given parser realizes. So, on the construal of "procedural rep-
resentation" as "representation of a procedure," there really can be some-
thing of substance to the distinction between procedural representations
and others.
But that's no use to Karmiloff-Smith. For there's no obvious reason that
parsing algorithms qua representations of procedures should be either
more or less accessible for the subject to report than grammars qua repre-
sentations of languages. Nor, as far as I can tell, is there any respectable
sense in which parsing algorithms are either more or less explicit than
grammars; they're just explicit about different things. So, for all I know, it
could be true that children start out having just a parsing algorithm and
end up having rules of grammar as well; and, for all I know, it could be
true that children are metalinguistically explicit about the rules of their
grammar but not about their parsing algorithm. But neither of these
claims would explain the other, so if both are true, then Karmiloff-Smith
has three worries on her hands rather than just the one she used to have.
There and Back Again 13 7
She used to have to explain why intramodular information gets more
accessible over time; now she has to explain that and also why the format
of intramodular information becomes less procedural over time, and why,
in general, procedural information should be supposed to be less acces-
sible than grammatical information.
To repeat: KarmiloH-Smith seems to think that if she could explain
why intramodular information becomes deproceduralized in the course of
development, that would explain why it becomes increasingly accessible
in the course of development. But it wouldn't. Here as elsewhere, format as
such is simply neutral with respect to accessibility as such.
There is, to be sure, another notion of procedural representation, which
I mention just to get it out of the way; it can't be what KarmiloH-Smith
has in mind. On this view, a procedural representation isn't a representa-
tion of anything; it's not even the representation of a procedure. To have
a procedural representation of how to solve a certain sort of problem is
just: to be able to solve the problem. In this (rather strained) sense, I have
access to a procedural solution of the problem of how to raise my arm;
that is, I am able to raise it. Maybe my ability to ride a bike is procedural
in that sense; I just can ride it, and no representations are invoked in
explaining how I do.
But, as I say, this can't be what KarmiloH-Smith has in mind since if
procedural representations aren't representations, then development can't
be a process of re-representation. Development can only change the
format of information that has a format; which, according to the present
construal, procedural representations do not.
2 "Redescription" as Inductive Generalization
Here's a second, and I think quite diHerent, kind of story that KarmiloH-
Smith tells about how it is that "representational redescription" accounts
for the demodularizatiqn of information in the course of cognitive
development-one that doesn't invoke the notion of procedural repre-
sentation. The child starts with representations of a variety of discrete
cases; subsequently, he generalizes over these, arriving at rules that (pur-
port to) apply to all the cases of a certain kind.
This is a kind of developmental progression that pretty likely does
occur; for example, it's presumably what generates the familiar phenom-
enon of "U-shaped" developmental curves. Here's a case that's well
known &om the literature on the child's developing mastery of past-tense
morphology. Ontogeny often exhibits the following phases:
4
phase 1: go/went
phase 2: go/goed
phase 3: go/went
138 Chapter 11
Apparently the child at phase 2 is operating on a generalization about the
past-tense which the child at phase 1 hasn't formulated: roughly, that you
make a past tense verb by sticking "-ed" on the end. In fact, at phase 2 the
child takes this generalization to be more reliable than it actually is;
he applies it indiscriminately to strong and weak verbs. By phase 3, he
has learned that "go" is an exception to the "add -ed" rule of past-tense
formation.
But though this sort of thing no doubt happens, it too is no help to
Karmiloff-Smith.
(i) Technical quibble Inductive generalization isn't really the "redescrip-
tion" of old information; it's the addition of new information. Universal
generalizations are stronger than the singular statements that instance
them. For one thing, they govern new cases. (If you know the "add-ed"
generalization, then you predict that "fricked" is the past tense of
"to frick." If you don't know the generalization, you can't make this
prediction.)
(ii) More important, there's every reason to think that this sort of "redescrip-
tion" is entirely intramodular. Adult speakers, by assumption, internally
represent the generalizations about their lanaguage of which children rep-
resent only the instances. But adults can't articulate such generalizations
any more than children can. (Thus most adults speakers know explicitly
that "walked" is the past tense of "walk" and that "vended" is the past
tense of "vend"; but they can't tell you why it is that in one case the
past-tense morpheme is sounded /t/ and in the other case it's sounded
/id/.) There is, in short, no reason to suppose that the kind of "rede-
scription" that goes on in formulating internal representations of such
inductive generalizations is a move in the direction of increased accessi-
bility. A fortiori, there's no reason to suppose that it's a move in the
direction of demodularization.
(iii) Redescription is like theory construction; the child is a "little theorist"
''Normally developing children not only become efficient users of lan-
guage; they also spontaneously become little grammarians" (32); "[they]
go beyond the input/output relations and reflect metalinguistically on the
word" (33 ). Likewise, "young children spontaneously come to theorize
about the physical world . . . by the internal process of representational
redescription which abstracts knowledge the child has already gained
from interacting with the environment" (78). And so forth in many places
in the book.
I'll be brief about this proposal. I have no doubt that (most, many, for
all I know, all) children reflect upon and theorize about the world; and I
have no doubt that their doing so comes as naturally to them as flight to
There and Back Again 139
butterflies. Nor do I doubt what Karmiloff-Smith is often at pains to insist
on, that what children do when they theorize is in many respects strik-
ingly similar to what grown-up scientists do when they are professionally
engaged; that, for example, there's a methodological premium on sim-
plicity, generality, and conservatism in both cases.
But there is, to my knowledge, no evidence at all for the claim that, in
children, theory construction is a process of demodualization. In par-
ticular, there's no evidence at all that the data that prompt the child to
theorizing are ever intramodular.
This connects with a point I made some way back. The natural assump-
tion is that what children theorize about is not what's represented in their
modules, but rather what's represented in the outputs that their modules
compute. According to this view, the child does have increasing reflective
access to something that she mentally represents; and, as her access to
what she mentally represents increases, so does her inclination to theorize
on the content of the representations. But what the child has increasing
access to is information in the structural descriptions that the modules
deliver, not intramodular information per se. Otherwise grownups could
introspect the grammars of their language and linguists would be out of
work. That would be a Very Bad Thing; my wife is a linguist, and I am
counting on her to support me in my retirement.
(iv) Connectionism I should add just a note about Karmiloff-Smith's sug-
gestion, in the last chapter of her book, that one might explicate the
notion of procedural {'1-level") representation by analogy to what goes
on in neural nets. As she reads the connectionist literature, a network rep-
resentation of the fad that both ''boy" and "girl" are English nouns con-
sists in the fad that there is a dimension of "the activation space [on
which] 'boy' and 'girl' in all their grammatical roles line up with all the
other words that we call nouns" (186). But this sort of representation of
their nounhood is (merely) implicit in the sense that "it is we, as external
theorists, who . . . label the trajectories through weight space as nouns,
verbs, subjects ... and so on. The network itself never goes beyond the
formation of the equivalent of stable Ievel-I representations .... The notion
of nounhood always remains implicit in the network's system dynamics.
The child's initial [sic] learning is like this, too. But children go on to
spontaneously redescribe their knowledge. The pervasive process of rep-
resentational redescription gives rise to the manipulability and flexibility
of the human representations" (186).
Now, I don't know how things work in children, but this does strike
me as a massive misdescription of how things work in networks (a topic
in respect of which, to be sure, massive misdescription has tended to
be the norm). In fad, the difference between classical and connectionist
140 Chapter 11
representations is orthogonal to the issue of explicitness. Here, for exam-
ple, is the pertinent fragment of a network that is fully explicit about
"boy" and "girl" both being nouns:
NOUN
1\
boy girl
Notice that this network is exactly as explicit about ''boy" and "girl"
being nouns as is the "classical" grammar of which the pertinent fragment
is:
Noun--+ boy
--+girl
In respect to explicitness, there is simply nothing to choose between the two
kinds or representation; the idea that classical architectures are somehow
ipso facto "explicit" or that connectionist architectures are somehow ipso
facto not explicit is simply confused.
I suspect that what causes the confusion is that the particular con-
nectionist model that Karmiloff-Smith has in mind takes for granted that
lexical categories like NOUN, ADJECTIVE, VERB, and the like are defined
distributionally. On that view, for a word to be a noun just is for it to
belong to a cluster of items with appropriately similar distributional
properties (that's what the stuff in the quotation about "similar trajectories
through state space" means). Correspondingly, a network whose connec-
tivity mirrors the relevant distributional patterns in its input can treat
NOUN and the like as "implicit" concepts in the sense that it treats them
as defined concepts. There's nothing more to being an (English) noun than
having much the same distributional properties (much the same trajectory
through state space) as paradigms like ''boy" or "apple."
But the question whether NOUN is in this sense an implicit concept is
simply orthogonal to the question of whether cognitive architecture is
connectionist. In particular, there is nothing about connectionist archi-
tecture per se that requires claiming that lexical classes have distributional
definitions, and there is nothing about the classical architecture per se that
requires denying that they do. That syntactic classes are distributionally
defined was, indeed, a main tenet of "taxonomic" linguistics; which was
classical from head to toe in its architectural assumptions.
5
I do want to be entirely clear about the polemical situation, if I can: I am
not saying that connectionist networks and classical grammars are other
than importantly different in their treatment of language. To the contrary;
there are crucial grammatical facts (particularly, facts about syntactic con-
stituency) that classical grammars capture naturally but that network
There and Back Again 141
grammars can't represent at all. (See chapters 8 and 9.) I do claim, how-
ever, that with respect to such facts as both kinds of models can represent,
there is no intrinsic difference in the explicitness with which they represent
them. In consequence, even if there were some ontogenetic process by
which children start out with connectionist networks in their heads and
end up with classical grammars in their heads, that wouldn't explain why
there is a corresponding shift from less to more explicit metalinguistic access.
To repeat the general moral of which the present case is an instance;
differences in format don't, in and of themselves, explain differences in
accessibility. So, even if there is an ontogenetic process of demodulariza-
tion (which I doubt), and even if there is an ontogenetic process of repre-
sentational redescription (which I also doubt), the redescription wouldn't
explain the demodularization.
Conclusion
Where does all this leave us? There are two issues that everybody is try-
ing to get clear on: What is the cognitive architecture of the adult mind?
and, whatever it is, how did it get to be that way? I think the empirical
results over the last couple of decades make a not implausible case that
modular architecture is an appropriate idealization for some aspects of the
organization of mature cognition. About the second question, however,
nothing is really known; we're all just playing our hunches. Perhaps the
deepest issue that divides people in the theory of cognitive development
is whether there are ontogenetic processes that affect cognitive architecture.
Karmiloff-Smith and Piaget are betting that there probably are; Chomsky
and I are betting that there probably aren't. All I can tell you for sure
is this: There may in .fact be architectural changes in the course of the
ontogeny of cognition, but nobody has found a clear case of one so far-
and since the developmental plasticity of the mind has been a main theme
of Anglo-American psychological speculation for a couple of centuries, I
do think that fact is striking.
Notes
1. I'm not as sure as I'd like to be about what an epigenetic process is. Karmiloff-Smith pro-
vides the following gloss on Piaget's usage: "For Piaget both gene expression and cogni-
tive development are emergent products of a self-organizing system that is directly
affected by its interaction with the environment" (9). But a self-organizing system that is,
nevertheless, "directly affected by its interaction with the environment" is, to my taste,
rather like a circle that nevertheless has comers. Probably "epigenetic" just means "not
primarily input driven." I shall assume that is what it means.
2. Even adults are more likely to overlook closed-class words than open-class words when
(e.g.) they are hunting for typos in a proofreading task; d. the relative inaccessibility of the
142 Chapter 11
cat" v . "John ~ d f d th
round th t adult do par "nt nc th"y
d in the tru tural de cription that par in
~ t l. 0 d ' I s m ('ffi h n ( I \ h I - r infl,rJTl
tion about -dosed-class i t ~ -that is in the sfrudur1" thscnpfions. Parity of argument sug-
gests the same explanation for the child's lack of metalinpstic access to information
about closed-class items.
3. The structural description of a passive doesn't, for example, teD you whether passivization
is a lexical transformation. Indeed, whether passive is a lexical transformation is just
the sort of question that linguists argue about. Clearly, linguists can't look inside their
modules either.
4. Earlier claims about just how prevalent this phenomenon is have recently come in for
reappraisal (see, e.g., Marcus et al., 1992). But the example will serve for expository pur-
poses.
S. The issue about whether there can be a connectionist account of language acquisition is
tacitly confounded with the issue of whether linguistic categories have distributional defi-
nitions not only in BM, but also in Elman et al., 1997. More on this in the next chapter.
Chapter 12
Review of Jeff Elman et al., Rethinking Innateness
Connectionism offers psychology a new kind of computational theory of
the cognitive mind. Unlike the familiar II classical" models of mental pro-
cesses, which borrow their architecture from digital computers, connec-
tionist networks run in parallel; they don't honor the distinction between
executive, memory, and program; and, although they are arrangements of
causally interacting representational states, they don't distinguish primi-
tive from complex representations and they don't operate by trans-
forming symbols. These differences are substantial, so there's a serious
argument going on. Readers who want an introduction to the connec-
tionist side could do worse than chapter 2 of Rethinking Innateness (here-
after RI); it provides a dear, capsule tutorial. (But acquaintance with the
classical approach to cognitive science is largely assumed, so RI isn't for
beginners.)
There is also an argument going on about what exactly it is that classi-
cal and connectionist theories are arguing about. Nativism, association-
ism, empiricism, rationalism, reductionism, genetics, computer science,
neuroscience, linguistics, ethology, and the mind/body problem (and I
may have forgotten a few) have all somehow involved themselves in
what started as just a psychologist's choice between theories of cognitive
architecture. This second argument is more fun than the first; it's what RI
is mostly concerned with; and it's what I focus on in this review. I'll start
with what RI takes (wrongly in my view) to be the core questions. Then
I'll say where I think the real issue lies. Innateness, as it turns out, is a bit
of a sideshow, but we'll get to it in due course.
Brainlikeness
Advertising for connectionist models of mind invariably emphasizes their
neurological plausibility. This line is developed in two ways io RI: On the
one hand, there's stress on analogies between the structure of computa-
tional networks and the structure of brain tissue. In this respect, the com-
parisons with classical cognitive theories are often invidious, not to say
sanctimonious (
11
[M]ost connectionist researchers are really committed to
144 Chapter 12
ultimate neural plausibility, which is more than you can say for most
other approaches" [SO]). On the other hand, the neurological data are said
to be that the human cortex is, ontogenetically, an "organ of plasticity."
That is supposed to be bad news for the view that innate "representa-
tional content" contributes to human cognitive development.
From the first of these claims, though it plays wonderfully in grant
proposals, it might be well to avert the gaze. Connectionist models are
networks of interconnected, unstructured objects called "nodes". Cortical
tissues are (among other things) networks of interconnected, but not
unstructured, neuronal cells. The nodes sort of look like neurons. From a
distance. And you can increase the resemblance by painting them grey.
Accordingly, the nodes "are likened to simple artificial neurons" (SO) ..
At this point, we could have a serious discussion of how good, or bad,
the analogy between nodes and neurons actually is. (The New York sub-
way system is also a network of connected nodes, and it too runs in par-
allel.) But that would be beside the point since it turns out, as one reads
further, that "the nodes in the models [are] equivalent not to single neu-
rons but to larger populations of cells. The nodes in these models are
functional units rather than anatomical units" (91). To what '1arger popula-
tions of cells" are the connectionist's nodes equivalent? Deafening silence.
There are exactly no proposals about how, in general, the putative func-
tional structures and the putative anatomical structures might relate to
one another. It's left wide open, for example, that nodes whose functions
are quite different might correspond to anatomically similar neural struc-
tures. Or that anatomical structures that are quite different might cor-
respond to functionally similar nodes. Or, indeed, that nodes might
correspond to brain stuff that doesn't count as a unit by any independent
criterion, anatomical or otherwise. All we're told is that the nodes are
"distributed" over the brain stuff. Somehow.
In fact it's entirely unclear how (or whether) connectionist cognitive
models are realized in actual brains, just as it is entirely unclear how (or
whether) classical cognitive models are realized in actual brains. Like
many other connectionist texts, RI changes its mind about this practically
from page to page. By the end, it comes down to the pious "hope ... that
such models will embody abstract and realistic principles of learning and
change that make contact with cognitive issues, while preserving some
degree of neural plausibility" (366). No doubt, "some degree of neural
plausibility" would be an excellent thing for a psychology to preserve; I'll
be sure to let you know if any of it should turn up.
The relation between evidence for cortical neuroplasticity and claims
about innateness is more interesting and more closely germane to the
special interests of Rl. The authors assume, literally without argument,
Review of Elman et al., Rethinking Innateness 145
that "[i]n the brain the most likely neural implementation for . . . innate
knowledge would have to be in the form of fine-grained patterns of syn-
aptical connectivity at the cortical level" (25). There is, however, a lot of
data suggesting that much of adult cortical structure emerges from com-
plex interactions between temporal, spatial, and neurochemical constraints
and is thus, presumably, not directly genetically encoded. There is plenty
of room to argue about the details; but, on balance, that's the current con-
sensus in brain science.
The appropriate inference is therefore that either there isn't much
genetically prespecifled information represented in the cortex, or that, if
there is, it isn't represented at the level of individual neurons and their
synaptic connections. The authors prefer (indeed, leap to) the former con-
clusion; but it's hard to see why. There are, after all, levels of organization
of the nervous system both larger and smaller than the interconnections
between individual neurons; and any of these other levels may perfectly
well be where innate-or, for that matter, learned-information is en-
coded. That one brick is a lot like another is not a reason to doubt that
the structure of brick buildings is, by and large, preplanned. All it shows is
that, if planning did go on, you have to look at aggregates of bricks to see
the effects.
The way things are set up in RI, an unwary reader might suppose it's
pretty clear, in general, how knowledge is encoded in the nervous system;
so if we don't find the right neural connectivity when we look at the
infant's brain, that shows that there's no innate content there. Nor, how-
ever, is there any learned content in the adult's brain if we insist on this
line of argument. For there isn't one, not one, instance where it's known
what pattern of neural connectivity realizes a certain cognitive content,
innate or learned, in either the infant's nervous system or the adult's. To
be sure, our brains must somehow register the contents of our mental
states. The trouble is: Nobody knows how-by what neurological
means-they do so. Nobody can look at the patterns of connectivity (or
at anything else) in a brain and 6gure out whether it belongs to some-
body who knows algebra, or who speaks English, or who believes that
Washington was the Father of his Country. By the same token, nobody
can look at an infant's brain and tell from the neurological evidence
whether it holds any or all of these beliefs innately. The sum and sub-
stance is that we would all like our cognitive models to be brainlike
because we all believe, more or less, that one thinks with one's brain. But
wishing won't make them so, and you can't make them so by proclama-
tion either. Nobody knows how to make a brainlike cognitive model,
classical or connectionist, because nobody knows how the cognitive brain
works.
146 Chapter 12
Interactions
'We argue throughout that an exclusive focus on either endogenous or
exogenous constraints fails to capture the rich and complex interactions
which occur when one considers organisms rooted in their environment
with certain predispositions, and whose interactions give rise to emergent
forms" (110); "our major thesis is that while biological forces play a cru-
cial role in determining our behavior, those forces cannot be understood
in isolation from the environment in which they are played out" (320).
But this can't really be RI' s major thesis; for who would wish to argue
against it7
For example: Linguistics is the locus classicus of recent nativist theoriz-
ing. But linguists might reasonably claim to be the only cognitive scien-
tists who have ever taken the interactionist program completely seriously.
Nobody doubts-linguists doubt least of all-that what language a child
learns depends on what language he hears; or that languages differ in lots
of ways; or that languages don't differ in arbitrary ways; or that any
normal child can learn any language of which he hears an appropriate
sample. The goal of linguistic inquiry, as classically conceived, is therefore
to provide a taxonomy of the range of variation between possible lan-
guages, and a precise account of what predispositions have to be available
to a creature that is able to learn any language within this range. What
more could an interactionist want in a research program?
The claim that behavior is shaped by the interaction of endogenous and
exogenous information isn't worth defending because it isn't in dispute.
Nobody has ever doubted it since (and including) Descartes. What's really
going on in RI is a certain view of what kinds of endogenous information
innateness can contribute to this interaction. "It is important to distin-
guish the mechanisms of innateness from the content of innateness .... We
suggest that for higher-level cognitive behavior, most domain-specific
outcomes are probably achieved by domain-independent means" (359);
'We discussed innateness in terms of mechanisms and content. . . . We
argue that representational [viz., content] nativism is rarely, if ever, a ten-
able position." (360-361).
Now we're getting closer.
Representational Nativism
Suppose it turned out (as, indeed, some of the data suggest it may) that
human infants act for all the world as though they expect unsupported
objects to fall. And suppose that the evidence was convincing that this
expectation couldn't have been learned. A natural hypothesis might then
be that the belief that unsupported objects fall is innate. RI wouldn't like
that at all, of course; but it's important to understand that it's not the
Review of Elman et al., Rethinking Innateness 147
innatetness per se that RI disapproves of. Officially all sorts of things can
be innate, for all that RI cares, so long as none of them is a belief; so long
as none of them has content.
Accordingly, much of RI is devoted to examples, sometimes ingenious,
sometimes even credible, where what look like the effects of innate content
can be explained as really the effects of innate mechanisms, typically in
interaction with one another and with environmental variables. Here is a
typical example: "why [do] beehive cells have hexagonal shapes .... [T]he
shape of the cell is the inevitable result of a packing problem: How to
pack the most honey into the hive space, minimizing the use of wax.
Information specifying the shape of the cell does not lie in the bee's head,
in any conventional sense; nor does it lie solely in the laws of physics
(and geometry). Rather, the shape of the cell arises out of the interaction
of these two factors" (320). Such interactions may occur at many levels of
scale, and local constraints don't need to do all the work. ''The outcomes
... typically involve complex interactions, including molecular and cellular
interactions at the lowest level, interactions between brain systems at the
intermediate level, and interactions between the whole individual and
the environment at the highest level" (320). Any of these interactions
may be biased by the genes, and to practically any degree. In short, RI will
tolerate as much innate control of ontogeny as you please-but no innate
content.
That's the official story. Actually, I doubt that mechanism vs. content
is the distinction that the authors really want. Imagine, for example, an
innate mechanism which, on activation (say by maturation or by some
environmental trigger), causes the child to believe that unsupported
objects fall. Presumably that would be just as objectionable as the innate-
ness of the belief itself. RI is pretty unclear about all this, but I suppose that
what it really wants to rule out is anything innate except general learning
mechanisms; specifically, general associative learning mechanisms.
But, anyhow, what's the objection to innate content? If representational
innateness is often the obvious theory of a creature's mental capacity, why
not suppose, at least some of the time, that that's because it's the right
explanation of the creature's mental capacity? It's not as though, in gen-
eral, there's a plausible alternative on offer; often there is none. Here, for
example, is what RI has to say about linguistic universals. "[Sometimes]
solutions are contained in the structure of the problem space. . . . The
grammars of natural language may be thought of as solutions of a reduc-
tion problem, in which multidimensional meanings must be mapped onto
a linear (one-dimensional) output channel" (386). But no examples are
given of how, even in sketch, an attested linguistic universal might be
explained this way; nor is anything said about why all sign languages,
which are spatially as well as temporally organized, obey much the same
148 Chapter 12
universal constraints that spoken languages do. And God only knows
what it means to say (or to deny) that meanings are "multidimensional."
In fact, as things now stand, there isn't any explanation of why there
are linguistic universals except the one that says they express innate rep-
resentational contents. Moreover, there are, in consequence of several
decades of investigation, detailed, abstract, and very powerful theories
about which ones the universal features of languages are, and how innate
knowledge of these features might interact with environmental data to
produce the adult speaker/hearer's behavioral phenotype. To be sure, as
RI keeps reminding us, these theories aren't "necessarily" true. But they
do make the handwaving about "reduction problems" look pretty feeble.
You might suppose, so vehement is RI' s rejection of content innateness,
that there is something in the architecture of connectionist models that
rules it out. Not at all. Assuming (what's far from obvious; but that's
another story) that connectionist networks can represent cognitive con-
tent at all, they can perfectly well represent innate cognitive content inter
alia. The authors themselves acknowledge this. So, the connectionism in
RI on the one hand, and RI' s rejection of content nativism on the other,
are quite independent. If somebody proved that content nativism is false,
that wouldn't tend in the least to show that connectionism is true; and if
somebody proved that content nativism is true, that wouldn't tend in the
least to show that connectionism is false. So, once again, what does RI
have against innate representational content? This is, I think, a real aporia.
You don't understand RI until you understand what's going on here.
Empiricism
'We are not empiricists," the authors of RI roundly claim (357). But all
this means is what we've already seen: RI tolerates innate mechanism but
jibs at innate content. By that criterion, the original empiricists weren't
empiricists either. Hume, for example, thought that the mechanisms of
association and abstraction are innate, and that the sensorium is too. All
he objected to was innate Ideas. In fact, like Hume, RI is into empiricism
up to its ears; it's the empiricism-not the neuroscience and not even the
connectionism-that calls all the shots. RI is less a connectionist per-
spective on development than an empiricist perspective on connection-
ism, its subtitle to the contrary notwithstanding.
There are, historically, two main ways people have thought about the
mind's commerce with experience. According to the picture that ration-
alists endorse (me too), cognitive processes are aimed at explaining experi-
ence. Typically, this requires solving problems of the form 'What must
the world be like, given that experience is such and sor' Such problems
engage the mind at every level, from the local to the cosmic. What must
the world be like such that these two visual points are seen reliably to
Review of Elman et al., Rethinking Innateness 149
move together? Answer: The points are on the surface of the same object.
What must the world be like such that the red shift of a star's light is
observed to correlate reliably with the distance to the star? Answer: The
universe is expanding. And likewise for problems of intermediate scale:
What must English be like such that people put the "John" before the
"runs" when they use it to say that John runs? Answer: in English, sub-
jects precede their verbs.
So, cognition exists to theorize about experience. One uses the con-
cepts at one's disposal to construct the theories; that's what concepts are
for. But where does one get the concepts? Some, no doubt, are boot-
strapped from the results of previous theorizing, but that process has to
stop somewhere (more precisely, it has to start somewhere). Hence the
traditional connection between rationalism and content nativism. Experi-
ence can't explain itself; eventually, some of the concepts that explaining
experience requires have to come from outside it. Eventually, some of
them have to be built in.
The empiricist picture is much different. According to empiricists, the
fundamental mental process is not theory construction but abstraction.
There is a certain amount of regularity in experience, and there is a certain
amount of noise. Statistical processes filter out mental representations of
the noise, and what's left are mental representations of the regularities. In
RI, as in much other empiricist theorizing, these statistical processes are
implemented by mechanisms for forming associative connections between
mental representations. Input nodes in connectionist networks respond
selectively to sensory properties of impinging stimuli; other nodes form
associative connections which, in effect, model correlations between the
responses of the input nodes. The picture, to repeat, is not significantly
different from the associative empiricism of Hume, though RI is rather
more explicit than Hume was about the nonlinearity of laws of association.
To have been anticipated by Hume is not a thing to be ashamed of,
to be sure. But one really should be aware of the enormous reductive
burden implicit in the empiricist program, and of how unlikely it is that
the required reductions could in fact be carried out. Arguably, you can
form the concept of red by abstracting it out of your experience because,
arguably, "red" is the name of a kind of experience. But you can't form the
concept of a chair that way because a chair isn't a kind of experience, cr a
logical construct out of experience, or a statistical function over experi-
ence. A chair is something mind independent, which, if it's in the vicinity
(and if the lights are on and one's eyes are open), causes one's experience
to pattern in the familiar way. Likewise, give or take a little, for nouns and
constellations. Empiricists were forever trying to reduce the concept of a
cause of experience to the concept of the experiences that it causes. No
wonder empiricism didn't work.
150 Chapter 12
Whether mental content reduces to experiential content is what RI is
really about. Sometimes, but only very rarely, RI notices this. For example,
in a study that has been influential in connectionist circles, Elman "trained"
a network to predict the next word in a sentence when given the preced-
ing fragment. In the course of discussing this network, RI remarks that:
''because both grammatical category and meaning are highly correlated [my
emphasis] with distributional properties, the network ended up learning
internal representations which reflected [my emphasis] the lexical category
of the structure of the words" (343 ). However, the problem of language
acquisition is that of how a child learns grammatical structure, not how
he learns correlates of grammatical structure. RI does remember this dis-
tinction for almost three pages. But then it quite forgets, and Elman's net-
work is described as having learned "distinctions between grammatical
categories; conditions under which number agreement obtained; differ-
ences between verb argument structure; and how to represent embedded
information" (345). These are not distributional correlates of grammatical
structures, notice; they're the very thing itself. The equivocation is strik-
ing; what explains it? I suppose RI must be assuming, tacitly, that gram-
matical structures really are just distributional patterns: To be a noun is
just to be (or to be distributionally similar to) a kind of word that appears
to the left of a word like "runs" (or to the left of words that are dis-
tributionally similar to "runs"); and so on.
The trouble with that sort of reductive proposal, however, is that the
reductions aren't ever forthcoming. In linguistics and elswhere, it invari-
ably turns out that there's more in the content of our concepts than there
is in the experiences that prompt us to form them. (For example: There
are, I suppose, nouns in Russian. If so, then being a noun couldn't be a
property that words have in virtue of their distribution in English.) That
there is generally more in the content of a concept than there is in the
experiences that prompt us to form it is the burden of the traditional
rationalist critique of empiricism. Friends more scholarly than I tell me
that Leibniz had that line of criticism pretty much worked out; Kant cer-
tainly did. It has never been seriously rebutted, though empiricists often
carry on as though it hadn't happened. As does RI, egregiously.
The empiricist program was in place for several hundred years; not just
in psychology but also-indeed especially-in epistemology and the
philosophy of science, where a lot of clever people wasted a lot of time
trying to reduce "theoretical" concepts (like TABLE) to "observation"
concepts (like RED and SQUARE). There were, in the course of all those
centuries, no successes at all; literally none; which is to say, not one. For the
sorts of reasons I mentioned above, pretty nearly everybody came to
conclude that the failures were principled. So it looked, until just recently,
Review of Elman et al., Rethinking 151
as though the argument between empiricism and rationalism had 6nally
been put to rest. The connectionists have revived it; but apparently with-
out quite realizing that that's what they're doing, and without, as far as I
can tell, having anything to add that changes the picture. So now I guess
we'll have to play out the argument between empiricism and rationalism
all over again. What Santayana said about history is also true of Philoso-
phy 101: Those who haven't understood it are destined to repeat it.
Chapter 13
Review of Steven Mithen' s The Prehistory of the
Mind
What's your favorite metaphor for minds? If you're an empiricist, or
ail associationist, or a connectionist, you probably favor webs, networks,
switchboards, or the sort of urban grid where the streets are equidistant
and meet at right angles: New York's Midtown, rather than its Greenwich
Village. Such images suggest a kind of mind every part of which is a lot
like every other. Take a sample here and you find the concept DOG asso-
ciated with the concepts CAT and ANIMAL; take a sample there and you
find the concepts ANTHONY and CLEOPATRA in close connedion.
Though mental content changes as you go from place to place in the
web, the structure is everywhere the same: It's concepts and connections
wherever you look.
But there is also an older, rationalist tradition of theorizing about the
mind, one that runs from Plato through Gall, Kant and the faculty psy-
chologists, to Freud and Chomsky. It too has its proprietary metaphors,
which are frequently architedural. The mind is like a building (Steven
Mithen thinks it's like a cathedral). Entrance and egress are variously con-
strained, and so too are the paths through the interior. There are public
places and private places, and places where the children aren't allowed to
go; and there are places in the attic that hardly anyone but Granny ever
visits. With that sort of arrangement, you might expect that information
would be quite unevenly diffused among the occupants; downstairs
knows things that upstairs doesn't, and upstairs doesn't know that down-
stairs knows them. Also, buildings are often divided into lots of little
spaces, with more or less specialized functions. These differences of func-
tion generally imply differences of structure, so you wouldn't expect to
learn much about how the kitchen is organized by examining the potting
shed. The long and short of the building metaphor is that the whole mind
doesn't always function as a system; and that when it does so, it's through
the interaction of its heterogeneous parts.
The empiricist idea of a homogeneous mind has dominated psycho-
logical and philosophical thinking in the English speaking countries since
the eighteenth century. But now the fashions seem to be changing, and
154 Chapter 13
the "modular" organization of cognition is widely promoted. Just what
this amounts to is far from clear, and some of the consensus is doubt-
less just in terminology. Still, people in a lot of fields of research have re-
cently come to think that some of our mental competences are isolated
and specialized in ways that the web/grid/network story doesn't easily
accommodate. There is even considerable agreement on which cognitive
capacities are the best candidates for the status of modules. The learning
and use of language is on everybody's list; Chomsky put it there, thus
initiating the new fashion in cognitive architecture. Commonsense biol-
ogy, commonsense physics, commonsense psychology, and aspects of
visual form perception are other favorite candidates. The bidding is still
open, but these five are the paradigms. If none of them is a mental
module, then probably nothing is.
Isolation and specialization are the defining properties of a module, so a
lot of the current discussion is about what precisely these consist in. A
word about this to set the stage for Mithen' s book.
First, specialization. Modules are supposed to be "domain-specific"
cognitive mechanisms. To say, for example, that there is a commonsense
psychology module is to claim that each of us has concepts and processes
available for inferring from the behavior of conspecifics to their states of
mind that are not likewise available for balancing the family checkbook
or deciding which omnibus to take to Clapham. Other people are a very
special part of the world; maybe one uses very special kinds of cognition
when one tries to understand them.
For instance, Brentano remarked on how odd it is that one can want
(need, seek, believe in) a gold mountain even though there aren't any
(whereas try climbing a gold mountain if there aren't any). Apparently,
the commonsense psychology module countenances relations between
minds and things-that-don't-exist. It may be that only phenomena in the
domain of the commonsense psychology module exhibit this logical
peculiarity; philosophers think so who hold that "intentionality" is the
mark of the mental. In similar spirit, Chomsky thinks that the grammatical
structures of all human languages are of the same formal type, and that
no communication system which failed to exemplify the universal format
could engage the operation of the language module. Are there likewise
eccentric constraints on the kinds of geometry that the commonsense
physics module can visualize? Does physical space have to strike us as
Euclidean even if it's not? Kant thought so.
What about the isolation of modules? Two kinds of ideas are current in
the cognitive science literature. First, modules are said to be "information-
ally encapsulated"; there are supposed to be architectural constraints on
the ways that information can flow between modules (and between them
and whatever unmodularized cognitive capacities the mind may have; see
Review of Mithen's The Prehistory of the Mind 155
below). The standard illustration is the persistence of perceptual illusions
even when one knows that they're illusory; you know perfectly well that
the moon is no bigger when it's on the horizon, but try convincing your
visual perception module that it doesn't look as though it is.
Second, modules are supposed to have proprietary developmental
careers. Language learning, for example, runs off largely independent of
the development of visual perception. Blind children are not, in general,
linguistically impaired; not even in their talk about space. Indeed, it's part
and parcel of the ontogenetic autonomy of the modules that their normal
development appears to make quite minimal demands on the character of
a child's environment. Children don't have to be taught their native lan-
guage; they just pick it up. Nor, apparently, do they have to be taught to
attribute people's behavior to the mental states they're in. Children tend,
quite naturally, to anthropomorphize whatever moves. What they have to
learn is which things don't have minds, not which things do.
Such considerations lead to the most spectacularly controversial thesis
that modularity theorists endorse: Modules are innate cognitive struc-
tures. They are, in Chomsky's phrase, "mental organs," and the autonomy
of their maturation is no more surprising than the maturational auton-
omy of, as it might be, legs from digestive systems. You wouldn't expect
fine-grained developmental measures to apply indifferently to knees and
stomachs, nor would you expect the development of either to depend on
special tuition. Rather, their maturation is internally driven and largely
mutually independent because it is genetically preprogrammed. So too
with mental modules; or so the story goes.
You might think that cognitive psychologists would reject this theory
out of hand for fear of becoming unemployed. If "cognitive development"
is just a cover term for the independent ontogenies of a heterogeneous
bundle of mental organs, what general laws of cognitive development
could there be for psychologists to discover? But this reaction is precip-
itous. It is, after all, not plausible that the mind could be made only of
modules; one does sometimes manage to balance one's checkbook, and
there can't be an innate, specialized intelligence for doing that. Maybe
cognitive development is the story of how the child somehow gets from
isolated, genetically preprogrammed, domain-specific cognitive modules
to the relatively unspecialized kind of intelligence that doing one's check-
book requires.
So, finally, to Mithen, who has been reading a lot of psychologists who
think that the child's cognitive development proceeds from a modularized
mind to one that's "cognitively fluid" (Mithen' s term). It's Mithen' s daring
and original suggestion that maybe this sort of ontogeny recapitulates
human phylogeny. Maybe the evolution from premodern man, to early
modem man, to us is the story of the emergence of cognitive fluidity
156 Chapter 13
from (or, rather, alongside of) the modularized components of our minds.
''The differences ... are analogous to those between Romanesque and the
succeeding Gothic cathedrals. . . . In GothiC architecture sound and light
can flow freely ... unimpeded by the thick walls and low vaults one Ands
in Romanesque architecture .... Similarly, in the phase 3 mental architec-
ture, thoughts and knowledge generated by specialized intelligences can
now flow freely around the mind .... " (71).
Well, it's a Ane idea, and a Ane book. At a minimum, Mithen' s easy
style is accessible to the casual reader who is looking for an overview, but
his book is also chock full of useful archaeological and anthropological
detail. A word to professionals: Do not skip the footnotes. They contain
stuff that will fascinate and edify the working cognitive scientist. Mithen
seduced me from the Anthony Powell that I was reading; I can't imagine
warmer commendation.
But is Mithen' s theory of cognitive phylogeny true? For that matter,
how well does it actually fit the ontogenetic theory that he uses as a
model? And, for another matter, is the ontogenetic theory that he uses as
a model itself true? Mithen is three levels deep in speculation: he needs
there to be mental modules; he needs the ontogeny of cognition to con-
sist in the maturation of these modules and, eventually, in the emergence
of cognitive fluidity; and he needs a phylogeny of cognition that recapit-
ulates this ontogeny. It's a lot to ask for, and I'm not convinced that it all
works.
By his own estimate, Mithen has two major epochs in the prehistory of
our minds to account for. There's the transition, ten million years or so
ago, from "the ancestral ape" (the presumed last common ancestor of us
and chimpanzees) to "early man"; and there's the transition, one-hundred
thousand years or so ago, from early man to "modem man" (in effect, to
us). This taxonomy is itself not untendentious. But Mithen Ands evidence,
from a variety of archeological and comparative sources, that each of
these major transitions implied a comprehensive increase of cognitive
capacity; so comprehensive as to suggest a corresponding reorganization
of the underlying cognitive architectures.
What sort of reorganization? Here's where Mithen wants phylogeny to
borrow from ontogeny. The ancestral ape had a mind that ran on "general
intelligence" together with, at most, a rudimentary modularized "social
intelligence"; whereas, in the next stage, the mind of early man had
modularized "technical intelligence" and "natural history intelligence" as
well. Finally, the emergence of cognitive fluidity from a relatively com-
partmentalized, modular mind is what Mithen thinks distinguishes early
man from us. It's this latter difference that Mithen likens to the change
from Romanesque to Gothic in the passage I quoted above.
Review of Mithen' s The Prehistory of the Mind 15 7
I do admire the attempt to find places for so many pieces. But it takes
some forcing to make them fit, and what Mithen ends up with shows the
signs of strain. For example: If cognitive ontogeny is indeed the develop-
ment of a fluid mind from a modular mind, then it offers a dear precedent
for the hypothesized shift from "early" to "modem" intelligence, much as
Mithen says. But what, exactly, is it in cognitive development that's sup-
posed to correspond to the emergence of early man's modular mind from
the rudimentary general intelligence of the ancestral ape7 Most cognitive
psychologists who are seriously into cognitive modules think that mod-
ules are what children start with; certainly that's the natural view if you
think that a modular intelligence is likely to be innate. And presumably it
has to be innate if, as Mithen supposes, modularity is a product of the
evolution of cognitive architecture.
Mithen relies heavily on some views of Karmiloff-Smith (1992; see
chapter 10), who does indeed envision three phases of ontogeny, including
both the assembly of rudimentary, task-specific capacities into modules,
and transitions from modular intelligence to something that's supposed to
be more fluid. But, precisely because she doesn't think modular archi-
tecture is genetically programmed, Karmiloff-Smith isn't by any means a
paradigm modularity theorist. In effect, Mithen is trying so hard to con-
struct an ontogenetic theory to ground phylogeny in, that he finds him-
self agreeing with people who don't really agree with one another-or
with him.
Still Karmiloff-Smith might be right that ontogeny doesn't actually start
with modules; or Mithen might be right about the prehistory of the mind
even if the psychologists are wrong about its development. More worry-
ing than whether Mithen can get phylogeny to run with ontogeny are the
internal tensions in his account of the former.
For example, for Mithen' s phylogenetic story to work, modules have to
be smarter than generalized intelligence in order to account for the tran-
sition from the ancestral ape to early man; and cognitive fluidity has to be
smarter than modules in order to account for the transition from early
man to us. But, in fact, it seems that Mithen takes "generalized intelli-
gence" and "cognitive fluidity" to be much the same sorts of things;
indeed, he sometimes describes the phylogeny of mind as an "oscillation"
from these to specialized intelligence and back again. But that can't be
right; the same cause can't explain opposite effects. There must be some-
thing very different between, on the one hand, the kind of general intelli-
gence that modular minds are supposed to have supplanted, and, on the
other hand, the kind of fluid cognition that Mithen thinks allowed us to
supplant Neanderthals. But Mithen doesn't say what this ingredient X
might be. Nor does he say how it could be that, whereas modules emerge
from general intelligence in the first transition, general intelligence
158 Chapter 13
emerges from modules in the second. Prima facie, one would think
"emerging from" ought to be a one way street.
The trouble-and it's not just Mithen' s trouble; it's a scandal in the
whole cognitive science literature-is that nobody knows what "general
intelligence" is or what "emerging from" is either. For example, there's a
strong inclination to say that minds that are specialized must somehow
evolve from minds that are less so; that sounds like good Darwinism. And
there's also a strong inclination to say that it's because our intelligence
is generalized and theirs is not that we're so much deverer than other
species. But, to put it mildly, it's not clear that you can have it both
ways. I suspect that cognitive scientists want "general intelligence" and
"emerging from" to paper over this crack, but that no coherent concepts
could. No wonder if nobody can figure out what general intelligence and
emerging from are.
Mithen does suggest that the transition from modularity to fluidity is
somehow mediated by the use of language. Here again, he is explicitly
echoing a literature in cognitive development according to which the
transition from modular to generalized intelligence in children is linguisti-
cally mediated. But neither Mithen nor anybody else has been able to
make clear how language manages to do the trick. The problem is that, if
specialized concepts are the only ones that you have, then words that
express specialized concepts are presumably the only ones that you can
learn. In effect, the linguistic mediation story replaces the question ''how
do you get from specialized concepts to generalized concepts" with the
question ''how do you get from specialized concepts to words for general-
ized concepts." There's nothing to indicate that the second question is
easier to answer than the first.
Well, any number can play this game. Here's my guess, for what it's
worth, at how the pieces fit together. The phylogenetically primitive form
of mind is not "generalized intelligence" but the reflex; that is, the more
or less instantaneous, unmediated elicitation of an organic response by an
impinging environmental stimulus. Two things have happened to reflexes
in the course of the phylogeny of mind, both tending to decouple the
stimulus from the response.
The first is the evolution of complicated, inferential processess of
stimulus analysis, which parse an organism's world into a wide variety of
behaviorally relevant categories; roughly, the higher the organism the
more such categories its perceptual repertoire acknowledges. Reflexes
respond to mechanical forces and to chemicals, light, gravity, and such; so
unless these sorts of stimuli signal an opportunity for behavior, a mind
that's made of reflexes is ipso facto blind to it. By contrast, our kinds of
minds envision a world of objects, persons, and relations between them,
Review of Mithen's The Prehistory of the Mind 159
upon the distribution of which the success of our behavior depends; so it's
into these categories that we want perception to articulate the world that
we act on. Phylogeny obliged, and apparently it started to do so early.
Mechanisms for high-level perceptual analysis are much older than any-
thing that could reasonably be called even rudimentary generalized intel-
ligence. For example, insects use such mechanisms when they navigate
(see Gallistel, 1990). Wherever they tum up, they exhibit the character-
istic properties of modules: computational isolation, specialization, and
genetic specificity. We probably have half a dozen or so of these modu-
larized perceptual analyzers, as I remarked above.
Modules function to present the world to thought under descriptions
that are germane to the success of behavior. But, of course, it's really the
thinking that makes our minds special. Somehow, given an appropriately
parsed perceptual environment, we often manage to figure out what's
going on and what we ought to do about it. It's a hobbyhorse of mine
that the thinking involved in this "figuring out" is really quite a different
kind of mental process than the stimulus analysis that modules perform;
and that it really is deeply mysterious. Suffice it that Mithen' s model of
phylogeny-that language somehow mediated a seepage of information
from module to module-surely underestimates the difference between
specialized intelligence and thought. Even if early man had modules
for "natural intelligence" and "technical intelligence," he couldn't have
become modem man just by adding what he knew about fires to what he
knew about cows. The trick is in thinking out what happens when you
put the two together; you get steak au poivre by integrating knowledge
bases, not by merely summing them.
But "integration" is just a word, and we haven't a smidgen of a theory
of what it is or how it happens.
On my view, the phylogeny of cognition is the interpolation, first of
modularized stimulus analysis and then of the mechanisms of thought,
between the Ss and Rs of aboriginal reflexive behavior. If that's right, then
perhaps cognitive phylogeny hasn't much to learn &om cognitive ontog-
eny after all. (Piaget taught that infants start out as sensory-motor reflex
machines and develop into formal intelligences, but practically nobody
outside Geneva believes that any more.) Quite possibly, the basic archi-
tecture of perception and thought are in place at birth, and what mostly
happens to the child's mind in development is what also mostly happens
to his body: Both get bigger.
I present this sketch as one kind of alternative to Mithen' s picture. But,
in fact, I wouldn't bet much on either. The synthesis of cognitive ontog-
eny with cognitive phylogeny is certainly premature; if somebody were
to tell us the right story right now, we probably wouldn't understand it.
160 Chapter 13
But t h ~ synthesis ~ to be premature if it's to be of any use to guide
research. We're not very good at thinking about minds; we need aU the
help we can get. Mithen has made a convincing case that "if you wish to
know about the mind, do not ask only psychologists and philosophers:
make sure you also ask an archaeologist." And he's drawn a 6rst map of
what might emerge &om putting aU their answers together. No doubt,
we'll have to do the geography again and again before we get it right; but
Mithen has made a start; more power to him.
Part IV
Philosophical Darwinism
Chapter 14
Review of Richard Dawkins's Climbing Mount
Improbable
"How do you get to Carnegie Hallr' "-Practice, practice.'' Here's a dif-
ferent way: Start anywhere you like and take a step at random. If it's a
step in the right direction, I'll say "warmer," in which case repeat the pro-
cess from your new position. If I say "colder," go back a step and repeat
from there. This is a kind of procedure that they call "hill climbing" in the
computer-learning trade (hence, I suppose, the title of Dawkins's book).
It's guaranteed to get you where you're going so long as the distance
between you and where you're going is finite (and so long as there are
no insurmountable obstacles or '1ocal maxima" in the way. Nothing is
perfect).
Hill climbing is often the theory of choice when a scientist's problem is
to explain how something got to somewhere you wouldn't otherwise
have expected it to be. That's in part because it's such an abstract and
general sort of theory. All it requires is a source of random variation, a
filter to select between the variants, and some "memory" mechanism to
ensure that the selected variations accumulate. In all other respects, you're
free to adapt it to whatever is worrying you.
For example, the '1tere" and "there" needn't be spatially defined. They
might be, respectively, the undifferentiated primal, protoplasmic slime and
the vast, intricate proliferation of species of organisms that now obtains.
Darwinism (or, anyhow, the adaptationist part of Darwinism) is a hill-
climbing account of the phenomenon of speciation: Genetic mutation
takes the place of your random milling about; the inherited genetic
endowments of successive generations of organisms correspond to the
succession of positions that you occupy between here and Carnegie; and,
instead of my shaping your path with gentle verbal cues, natural selection
determines the direction of evolution by killing off mutations that happen
to reduce organic fitness.
It all sounds pretty plausible. It might even be true. But the fact that a
hill-climbing model could, in mathematical principle, find a way from
where you started off to where you ended up, doesn't at all imply that
you (or your species) actually got there that way. I could have just told
164 Chapter 14
you where Carnegie Hall is, and you could have gotten there by follow-
ing my instructions ("directed evolution"). Or I could have picked you
up and carried you ("interventionism"). Or you could have started out
at Carnegie, in which case the question wouldn't have arisen ("pre-
formation"). No doubt, other possibilities will occur to you. In the present
case, one might reasonably wonder whether we did, after all, get to be us
in the way that Darwinian adaptationism says that we did; it's reasonable
to want to see the evidence.
Especially so because the scientific success of the hill-climbing style of
explanation has often been underwhelming in other areas where it has
been tried. Classical economics (by which Darwin was apparently much
influenced) wanted to use it to account for the organization of markets. In
a system of exchange where gizmos are produced with randomly differing
efficiencies, canny consumers will filter for the gizmos that are best and
cheapest. Gizmos that are too expensive to buy or too cheap to sell at a
profit will automatically be screened out. Eventually an equilibrium will be
achieved that comports, as well as can be, with all the interests involved.
That's a nice story too. But, in the event, what often happens is that
the big gizmo-makers buy out the little gizmo-makers and supress their
patents. If there's still more than one gizmo-maker left in the field, they
compete marginally by painting their gizmos bright colors, or paying
some airhead to praise them on television. The evolution of gizmos
therefore grinds to a halt. Whichever producer a consumer decides to buy
his gizmos from, he finds that they don't work, or don't last, or cost too
much.
For another example, consider a version of hill-climbing theory that
used to be popular in psychology. How does a creature's behavior get
organized? How, for example, do you get from being a babbling baby to
being a fluent speaker of English? Here's how, according to B. F. Skinner
and the tradition of "reinforcement theory": Babbling is vocal behavior
that's produced at random. When you happen to make a noise that
sounds sort of like the local dialect, "society" reinforces you; and your
propensity to make that sort of sound (or better, your propensity to make
that sort of sound in those sorts of circumstances) increases correspond-
ingly. Keep it up and soon you'll be able to say "Carnegie Hall" or "Jasha
Heifitz" or any other of the innumerable things that being able to speak
English allows you to say. Skinner used to complain to people who didn't
like his story about learning that he was just doing for the formation of
behavior what Darwin did for the formation of species. There was, I think,
some justice in that complaint, but it's an argument that cuts both ways.
In the event, language learning doesn't work by Skinnerian hill climb-
ing. In particular, language learners don't make their errors at random in
the course of the acquisition process. Rather, as Noam Chomsky famously
Review of Dawkins's Climbing MounllmproWI4 165
pointed out, the grammatical and phonological hypotheses about lan-
guage structure- that children ever think to try out are sharply endoge-
nously constrained. 'Who Mummy lover' is recognizably baby talk; but
'1ove Mummy whor' is not; it just isn't the kind of thing children say in
the course of acquiring English. Ergo, it's not the kind of thing that
society is ever required to filter out in the course of "shaping" the child's
verbal behavior. But why isn't it, if children are hill climbing toward the
mastery of English grammar, making mistakes at random as they go7
So there are at least two cases where, pretty clearly, applications of hill-
climbing models tell less than all there is to be told about how a system
gets organized. These examples have something striking in common. Hill
climbing wants a random source of candidates to filter; but, in the market
case and the language acquisition case, it appears that there are "hidden
constraints" on which candidates for filtering ever get proposed. The
market doesn't produce its gizmos at random, and the child doesn't
produce its verbalizations at random either. The market is inhibited by
restraint of trade, the child by (quite possibly innate) conditions on the
kinds of language that human beings are able to learn and use. No doubt,
in both cases, there is some residual random variability, and correspond-
ingly some filtering which serves to smooth rough edges; so hill climbing
gets a sort of vindication to that extent. But it's pyrrhic if, as practicioners
in economics and psycholingustics tend to suppose these days, it's the
hidden constraints that are doing most of the work.
So the track record of hill-climbing explanations outside biology isn't
what you'd call impeccable. What, then, about speciation? Nobody with
any sense doubts that adaptation is part of the truth about evolution; but
are there, maybe, ''hidden constraints" at work here too7 Or is the envi-
ronmental filtering of random mutation most of what there is to how
creatures evolve? Nobody loses absolutely all of the time. Maybe speci-
ation is where hill climbing wins.
There is, in fact, currently something of a storm over just this issue, the
vehemence of which Dawkins's book is much too inclined to understate.
Paleontologists, since Darwin's own time, have often complained about
what looks, from an anti-adaptationist perspective, like an embarassing
lack of smooth gradations from species to species in the geological record.
Maybe evolution gets from place to place by relatively big jumps ("sal-
tations''), the intermediate options being ruled out by hidden contraints
on what biological forms are possible. Something like this idea is at the
heart of the current enthusiasm for evolution by "punctate equilibria." If
you want to get to Carnegie, don't bother with exploring the inter-
mediate loci; take a jet.
Dawkins doesn't make much of this sort of option; he's too busy
assuring his lay audience that everything is perfectly fine chez classical
166 Chapter 14
adaptationism. Issues about evolution have become so politicized that a
popularizing biologist must be tempted to make a policy of pas devant les
enfants. Dawkins has, I think, succumbed a bit to this temptation. It's a
disservice to the reader, who thereby misses much of the fun. For a cor-
rective, try Niles Eldredge's 1995 book, Reinventing Danoin.
Anyhow, if classical adaptationism is true, then, at a minimum, the
route from species A to its successor species B must be composed of
viable intermediate forms which are of generally increasing fitness; there
must be, in Dawkins's metaphor, smooth gradients leading up the hill that
adaptation climbs. Much of Dawkins's book is devoted to an (admirable)
attempt to make the case that there could have been such viable inter-
mediates in the evolution of vision and of winged flight. Dawkins doesn't
(and shouldn't) claim that any of these intermediate creatures are known
to have existed. But he is pretty convincing that they might have existed,
for all that biochemistry, physiology, embryology, and computer model-
ing have to tell us. The naive objection to adaptationism is that random
mutation couldn't have made anything as intricate as an eye. Dawkins's
answer is that sure it could; there's a physiologically possible path from
bare sensitivity to light to the kind of visual system that we've got, and
overall fitness would plausibly increase and accumulate as evolution tra-
verses the path. It appears, in fact, that there may be several such paths;
eyes have been independently reinvented many times in the course of
evolution.
It is, however, one thing to show that evolution might have been
mostly adapatation; it's another thing to show that it actually was. Many
readers may be disappointed that Dawkins doesn't discuss at all the evo-
lution of the piece of biology that they are likely to have most at heart:
namely, human cognitive capacities. This is, on anybody's story, one of
the places where the apparent lack of intermediate forms seems most
glaring. Cognition is too soft to leave a paleontological record. And, pace
some sentimental propaganda on behalf of chimpanzees and dolphins,
there aren't any types of creatures currently around whose cognitive
capacities look even remotely similar to ours. Moreover, there is a prima
facie plausible argument that hidden constraints might play a special role
in the evolution of a creature's psychological traits as opposed to, say, the
evolution of its bodily form.
It's truistic that natural selection acts to filter genetic variation only
insofar as the latter is expressed by corresponding alterations of a
creature's relatively large-scale structure (alterations, for example, of the
organs that mediate its internal economy or its environmental inter-
actions). The slogan is: Genetic variants are selected for their phenotypic
fitness. This holds, of course, for the case of nervous systems too; genetic
endowments build neurological structures which natural selection accepts
Review of Dawkins's Climbing Mount Improbable 167
or rejects as it sees fit. Suppose that there is indeed relatively un-
systematic variation not only in the genetic determinants of neurological
structure, but also in the corresponding neurological phenotypes. Still,
brain structures themselves are selected for the fitness of the psycho-
logical capacities that they support. They're selected, one might say, not
for their form but for their function. And nothing generai-I mean
nothing general-is known about the processess by which neurological
alterations can occasion changes in psychological capacities.
Gradually lengthening the giraffe's neck should gradually increase its
reach; that seems sufficiently transparent. But it's wide open what tumul-
tuous saltations a gradual increase in (as it might be) brain size or the
density of neural connections might cause in the evolving cognitive
capacities of a species. The upshot is that even if we knew for sure that both
genetic endowments and neurological phenotypes vary in a way that is
more or less random and incremental, just as adaptationism requires, it
wouldn't begin to follow that the variation of psychological traits or
capacities is random and incremental too. As things now stand, it's per-
fectly possible that unsystematic genetic variation results in correspond-
ingly unsystematic alteration of neurophysiological phenotypes; but that
the consequent psychological effects are neither incremental nor con-
tinuous. For all anybody knows, our minds could have gotten here largely
at a leap even if our brains did not. In fact, insofar as there is any evidence
at all, it seems to suggest that reading brain structures onto mental capac-
ities must substantially amplify neurological discontinuties. Our brains
are, by any gross measure, physiologically quite similar to those of crea-
tures whose minds are nonetheless, by any gross measure, unimaginably
less intelligent.
Dawkins likes to "insist ... that wherever in nature there is a sufficiently
powerful illusion of good design for some purpose, natural selection is the
only known mechanism that can account for it" (202). He's right, I think,
but this is another of those double-edged swords. The conclusion might
be that adaptation really is most or all of what there is to evolution; or it
might be that we don't actually know a lot about the etiology of what
appears to be good biological design. Dawkins is inclined to bet on the
first horse, but it's not hard to find quite reputable scientists who are
inclined to bet on the second. Either way, it's a shame not to tell the
reader that what's going on is, in fact, a horse race and not a triumphal
procession.
Dawkins is the kind of scientist who disapproves of philosophy but
can't stop himself from trying to do some. That's quite a familiar syn-
drome. I should say a few words about what I'm afraid he takes to be the
philosophical chapters of his book. They are, in my view, a lot less inter-
esting than the biology.
16& Chapter 14
Dawkins says, rightly, that Darwinism teaches us that the biological
population of the world wasn't made for our comfort, amusement, or edi-
fication. 'We need, for purposes of scientific undentanding, to find a less
human-centered view of the natural world." Right on. But then he spoils it
by asking, in effect, if it's not all for us, who (or what) is it all for? This is a
bad question, to which only bad answers are forthcoming.
The bad answer Dawkins offers in the present book follows the same
line that he took in The Selfish Gene: It's all in aid of the DNA. "[W]hat are
living things really [sic] for .... The answer is DNA. It is a profound and
precise answer and the argument is watertight .... " (24 7). The idea is that,
from the gene's point of view, organisms are just "survival machines"
whose purpose is to house and propogate the DNA that shaped them. A
creature's only function in life (or in death, for that matter; see Dawkins's
adaptationist treatment of the evolution of altruism) is to mediate the
proliferation, down through the generations, of the genes that it carries.
Likewise for the parts of creatures: ''The peacock's beak, by picking up
food that keeps the peacock alive, is a tool for indirectly spreading instruc-
tions for making peacock beaks" (252) (i.e., for spreading the peacock's
DNA). It is, according to Dawkins, the preservation of the genetic instruc-
tions themselves that is the point of the operation.
But that doesn't work, since you could tell the story just as well from
the point of view of any other of a creature's heritable traits; there's noth-
ing special, in this respect, about its genetic endowment. For example,
here is the Cycle of Generation as it appears from the point of view of the
peacock's selfish beak: "Maybe genes think what beaks are for is to help
make more genes, but what do they know about philosophy? Beaks see
life steadily and they see life whole, and they think what genes are for is to
help make more beaks. The apparatus-a survival machine, if that amuses
you-works like this: beaks help to ensure the proliferation of peacocks,
which help to ensure the proliferation of peacock DNA, which helps to
ensure the proliferation of instructions to make more peacock beaks, which
helps to make more peacock beak.t;. The beaks are the point; the beaks are
what it's all'for.' The rest is just mechanics."
What's wrong with this nonsense is that peacock beaks don't have
points of view (or wants, or preferences), selfish or otherwise. And genes
don't either (not even "unconsciously" though -Dawkins is often confused
between denying that evolutionary design is literally conscious and deny-
ing that it is literally design. It's the latter that's the issue). All that happens
is this: microscopic variations cause macroscopic effects, as an indirect
consequence of which sometimes the variants proliferate and sometimes
they don't. That's all there is; there's a lot of "because" out there, but
there isn't any "for."
Review of Dawkins's Climbing Mount Improbable 169
In a certain sense, none of the teleological fooling aroimd actually
matters (which is, I guess, why Dawkins is prepared to indulge in it so
&eely). When you actually start to do the science, the metaphors drop
out and statistics take over. So I wouldn't fuss about it except that, like
Dawkins, I take science philosophically seriously; good science is as close
as we ever get to the literal truth about how things are. I'm displeased
with Dawkins's pop gloss on evolutionary theory because I think it gets
in the way of seeing how science shows the world to be; and that, I would
have thought, is what the popularizer of science-as-philosophy should
most seek to convey. Dawkins is rather proud of his hardheadedness (he
writes "sensitive" in sneer-quotes to show how tough he is); but in fad I
think his naturalism doesn't go deep enough. Certainly it doesn't go as
deep as Darwin's.
It's very hard to get this right because our penchant for teleology-
for explaining things on the model of agents with beliefs, goals, and
desires-is inveterate, and probably itself innate. We are forever wanting
to know what things are for, and we don't like having to take Nothing for
an answer. That gives us a wonderful head start on understanding the
practical psychology of ourselves and our conspecifics; but it is one of the
(no doubt many) respects in which we aren't kinds of creatures ideally
equipped for doing natural science. Still I think that sometimes, out of the
comer of an eye, "at the moment which is not of action or inaction," one
can glimpse the true scientific vision; austere, tragic, alienated, and very
beautiful. A world that isn't for anything; a world that is just there.
Chapter 15
Deconstructing Dennett's Darwin
Nobody likes me,
Everybody hates me,
I'm going to go out in the back yard and
Dig up some worms and
Eat them.
-Song we used to sing around the campfire when we were very young
Introduction
Dan Dennett's new book, Darwin's Dangerous Idea,
1
is full of anecdotes,
digressions, thought experiments, homilies, "nifty" examples, and other
erudite displays; with such a density of trees per unit forest, it's easy to
lose one's way. But I think there is discernibly a main line to Dennett's
argument. I think it goes like this: Dennett offers a sketch of a metaphy-
sical construction in which the (neo-)Darwinian account of evolutionary
adaptation is supposed to ground
2
a theory of natural teleology; and this
theory of natural teleology is in tum supposed to ground an account of
the meaning of symbols and of the intentionality of minds. This program
is, of course, enormously ambitious; much more so than adaptationism per
se. So you can without inconsistency buy the Darwin stuff but not the
teleology; or you can buy the adaptationism and the teleology but not
the theory of intentionality; or you can buy all three but doubt that
Dennett has got it right about how they relate to one another.
My own view is that adaptationism probably isn't true; and that, even
if it is true, there probably isn't any notion of natural teleology worth
having; and that, even if adaptationism is true and there is a notion of
natural teleology worth having, the latter isn't grounded in the former;
and that even if adaptation grounds a theory of natural teleology, natural
teleology has nothing much to do with the metaphysics of meaning.
It might be fun to explore all of that, but life is short and I have tickets
for the opera. So, here's what I propose to do instead. I'll say just a little,
mostly pro forma, about natural selection and natural teleology and the
172 Chapter 15
connections between them; then I'll suppress my qualms and grant, for the
sake of argument, a sense of selected for and of function in which adapta-
tions can be said to be selected for the function that they perform. (I think
that what I'm thus proposing to concede, though it's pretty thin, is all the
natural teleology that Dennett actually uses in his account of intention-
ality.) The main discussion will then concentrate on what I take Dennett
to take to be the metaphysical connection between the teleological char-
acter of natural selection and the intentional character of thought. I'll
argue, contra Dennett, that there isn't any such connection.
Adaptation
Dennett sometimes comes on pretty strong about the central status of the
adaptationist paradigm in the overall scientific world view. "Adaptationist
reasoning is not optional: it is the heart and soul of evolutionary biology.
Although it may be supplemented, and its flaws repaired, to think of dis-
placing it &om central position in biology is to imagine not just the
downfall of Darwinism but the collapse of modem biochemistry and all
the life sciences and medicine" (238). That sounds pretty fierce, alright; if,
like me, you're inclined to think that just about everything in this part of
the woods is "optional," it should scare you half to death.
But it isn't clear that Dennett really means it. At one point, for example,
he contemplates the possibility that Gould is right and there are "hidden
constraints" on genotypic variability (and/or its phenotypic expression)
which may substantially reduce the space of options that natural selection
ever gets to choose &om: ''The constraints of inherited form and devel-
opmental pathways may so channel any change that even though selec-
tion induces motion down permitted paths, the channel itself represents
the primary determinant of evolutionary direction" (Gould, quoted by
Dennett, 25 7). You might think that Dennett would reply that this would
make the sky fall; but, in fact, he doesn't. His rejoinder is merely meth-
odological: "[Even if) hidden constraints guarantee that there is a largely
invisible set of maze walls ... in the space of apparent possibility ... we
still can do no better in our exploration of this possibility than to play out
our reverse-engineering strategies at every opportunity ... " (261). Or, as
he puts it in a summary passage, "Good adaptationist thinking is always
on the lookout for hidden constraints, and in fact is the best method for
uncovering them" (261). That makes it look as though there is, after all,
practically nothing that an adaptationist is required to believe about how
evolution actually works; he's only required to buy into a methodological
claim about how best to find out how it does. (Dennett doesn't, by the
way, give any argument that assuming adaptationism is the "the best
method" for uncovering hidden constraints; or even that it's a good
Deconstructing Dennett's Darwin 173
method-as compared, e.g., with the direct investigation of the bio-
physics of genotypic variability. I don't blame him; it's notoriously hard
to predict which way of doing a piece of sdence will work. The posi-
tivists were right about there not being a logic of discovery.)
So it's a bit unclear just what degree of commitment to adaptationism
Dennett requires of biologists who wish to be in good sdentific repute.
It's also unclear exactly what doctrine it is that he requires them to
endorse. Sometimes it seems pretty tepid: However it was we got
here from the primal protoplasmic slime, no miracles (" skyhooks") were
required on route. One might indeed take a hard line on that-I am
myself prepared to sign on-since it is asymptotically close to empty. If
we did come across a miracle, we'd call it a basic law or an anomaly and
swallow it down accordingly. (Cf. eledridty, action at a distance, and
what g ~ on inside black holes.) Unlike many of my philosophical col-
leagues, I do think that there are substantial constraints that physicalism
and reductionism impose on how we choose between special sdence
theories. But these constraints are subtle and abstract. It's unimaginable
that they should have the force of an a priori endorsement (or prohibition)
of any substantive evolutionary theory.
Sometimes, however, what Dennett means by adaptationism isn't
merely methodological and really does have teeth; it includes not just the
familiar Darwinian schema incremental, random variability -+ natural selec-
tion for fitness -+ inheritance, but such very tendentious views as Dawkins's
"selfish geneH story, according to which the Darwinian schema has pri-
mary applicability at the level of the individual gene (as opposed to, say,
the unit of selection being the genotype, or the organism, or the breeding
group, or the species). It's hard to imagine that Dennett really thinks that
that theory is other than optional, or that the integrity of biology, or "the
life sciences and medidne," or of "secular humanism" (see p. 476) turns
upon it. But so his text sometimes suggests.
I propose not to get involved in this; it will do, for my purposes, merely
to stipulate. So, let an adaptationist be, at a minimum, somebody who
thinks that the explanation of many interesting phenotypic properties-
and, in particular, of many of such phenotypic properties as distinguish
spedes-invokes a history in which:
1. there is substantial and largely unsystematic genetic variation;
2. in consequence of which, there is substantial and largely unsys-
tematic variation in the corresponding phenotypes;
3. natural selection operates differentially upon the phenotypic
variants depending on their relative "fitness"; and
4. there is correspondingly differential inheritance of the genotypes
of the selected forms.
174 Chapter 15
I suppose something like this to be about as relaxed as (neo-)Darwinism
can g ~ t short of lapsing into vacuity. For example, this formulation is
neutral about what "fitness" consists in; it might be likelihood of individ-
ual survival, or likelihood of group survival, or likelihood of mating, or of
contributing to the gene pool, or whatever. Also, the doctrine I've ges-
tured toward isn't, as it were, "principled." One could hold that some-
thing like what it calls for really does happen some of the time, but that it
doesn't happen all of the time; or that it's a part, but not the only part, of
how evolution works; or that, though it's often true, it's wrong about the
evolution of some interesting phenotypic properties of, for example, us.
And so on.
I think that Dennett considerably overestimates the extent to which
there is currently an adaptationist consensus in evolutionary theory (see
the review by H. A. Orr, 1996 ). My own view is that Gould, Lew on tin,
Eldredge, and that crowd have made a substantial case that even this sort
of relaxed Darwinism is right about a good deal less than the whole
nature of evolution. I don't, however, propose to argue that this is so. For
one thing, I have no claim at all to expertise in this area; and, for another,
there's every indication that those guys can fight their own battles (see
especially Eldredge, 1995). I can't, however, resist a word on why I'm
particularly dubious about the application of even a relaxed Darwinism to
the phenotypic properties that I care about most, namely, human cogni-
tive traits and capacities.
Suppose that there is indeed a lot of relatively unsystematic genetic
variability. It's truistic (and also not in dispute) that natural selection can
ad to filter this variability only insofar as it is expressed by correspond-
ing phenotypic variation; genetic variants are selected for their phenotypic
"fitness." Now, there are all sorts of ways in which genetic variation
might tum out to be invisible at whatever level natural selection works on
(individual phenotype; species phenotype; or what have you). And there
are likewise all sorts of ways in which the magnitude of a genetic varia-
tion might underestimate the consequent alteration of a phenotype. For
one thing, as Dennett rightly stresses, the phenotypic expression of a
genotypic change is mediated by processes of (e.g.) protein synthesis that
are, to put it mildly, complicated. It's thus open in principle, and also in
empirical fad, how much genetic variability is dampened or amplified on
the route to building an organism or a species: It can tum out that many
genotypes converge on much the same phenotypes; or that only slightly
diHerent genotypes get grossly diHerent phenotypic expression; or that
identical genotypes get diHerent phenotypic expressions in diHerent con-
texts; etc. In short, it's perfectly possible that even if genetic variation is
unsystematic (no "hidden constraints" at the genetic level) the processes
that "read" it as phenotypic variation aren't. Thus Eldredge (op. cit.)
Deconstructing Dennett's Darwin 175
remarks that "[t]he implacable stability of species in the face of [underlying]
genetic ferment is a marvelous demonstration that large-scale systems
exhibit behaviors that do not mirror exactly the events and processes
taking place among their parts .... " (175)
All this holds, of course, for the psychological case inter alia; except
that here it's one 6llip worse since there is still another locus at which
''hidden constraints" may have their effects. This point isn't novel, but it's
well worth insisting on.
Suppose that there is relatively unsystematic variation in whatever are
the genetic determinants of neurophysiological structure, and suppose
further that this relatively unsystematic genetic variation is faithfully
expressed by a correspondingly unsystematic variation of neurophy-
siological phenotypes. Still, the selection of these neurophysiological
phenotypes must itself be mediated by whatever processes govern their
expression as psychological traits and capacities.
3
And nothing general-
! mean nothing general-is known about these processes. The upshot is
that even if we knew for sure that both genetic and neurophysiological
variability are more or less incremental, random, and unsystematic, it
wouldn't begin to follow that the variability of psychological traits or
capacities is too.
So, for example, as Chomsky and Gould have frequently remarked, it is
perfectly possible that there is random variability of, and environmental
selection for, overall brain size, and that this is as gradualistic and classi-
cally Darwinian as you like. That would still leave it open that, at some
point(s) or other, and by processes currently utterly unknown, the adven-
titious consequences of such physiological change are radical discontinuities
in behavioral repertoires. (Or, if you don't like brain size, try brain weight,
or brain/body ratio, or the density of neural connections, or the amount
of surface folding, or the proportion of cortex to old brain, or the pro-
portion of neurons to glia or, for that matter, the color of the damned
thing. Take your pick. All except the last have been in fashion, at one time
or another, as candidates for conveying selectional advantage.) I repeat,
nothing general is known about how physiological variation determines
variation of psychological traits and capacities. Do not believe anything
to the contrary that you may have read in the Tuesday New York Times.
To summarize this line of thought: It is, of course, an empirical issue
how close to being true even a relaxed adaptationism will prove to be.
For it to be very close, the effects of hidden constraints on the course of
evolution have to be relatively small compared to the effects of unsys-
tematic variation of genotypes and their phenotypic expressions. In the
psychological case, however, the notion of "phenotypic expression" is
ambiguous; it may refer to a creature's neurophysiology or to its behav-
ioral traits and capacities. So there are not just one but two places where
176 Chapter 15
hidden constraints may enter into determining the evolutionary process.
Even if the physiological expression of genotypic variation is largely uncon-
strained, there is, as things now stand, no reason to suppose that the
behavioral expression of physiological variation is too. In fact, insofar as there
is any evidence at all, it seems to suggest that reading neural structures
onto behavioral repertoires must substantially amplify neurophysiological
discontinuities. Our brains are, by any gross measure, physiologically
quite similar to those of creatures whose psychological capacities are
nonetheless, by any gross measure, unimaginably inferior.
Pace Dennett, even a relaxed adaptationism about our psychological
traits and capacities isn't an article of scientific faith or dogma; we'll just
have to wait and see how, and whether, our minds evolved. As of this
writing, the data aren't in.
Adaptation and Teleology
Suppose, however, that adaptationism is true; is it able to ground a notion
of natural teleology? Or, to put it in terms a little closer to Dennett's,
suppose there is a type of organism 0 that has some genotypic property
G that was selected because G's (characteristic) phenotypic expression P
increases O's relative fitness (on average, ceteris paribus, in ecologically
normal circumstances ... blah blah, blah). Is it then reasonable to speak of
Pas a property that 0 was "designed" to have? Or as a "solution" to an
"engineering problem" that 0' s ecology posed?
It's important to see how this kind of question might bear on issues
about intentionality. If adaptations are, in some not entirely figurative
sense, design features of organisms, then maybe the root metaphysical
problem about how to get meaning into a purely mechanical world order
is on its way to being solved. Just how this might go will concern us
presently. It will do, for now, to notice that DESIGNED FOR F-ING, and
the like, belong to the same family of concepts as DESIGNED WITH
F-ING IN MIND. Perhaps, then, the putative connection between natural
selection and design is the thin end of the intentional iceberg. Perhaps
there is a route from the indubitably mechanistic adaptationist story about
how phenotypes evolve, to the intentional idiom which (eliminativists
excepted) everybody thinks that a good theory of the mind requires.
That, in a nutshell, is how Dennett is hoping that things will pan out.
The subtext is the thing to keep your eye on here. It is, no doubt, an
interesting question in its own right whether adaptationism licenses teleo-
logical notions like SELECTION FOR. But what makes that question
interesting in the present metaphysical context is that SELECTION FOR
is presumably intensional.
4
Just as you can believe that P and not believe
that Q even though P iff Q, so a creature might be selected for being F
Deconstructing Dennett's Darwin 177
and not for being G even though all and only the Fs are Gs. Correspond-
ingly, the issue we're most concerned with is whether a naturalistically
grounded notion of selection for would be intensional in the ways, and for
the reasons, that mental states are. If so, then maybe a naturalistic teleol-
ogy is indeed a first step toward a naturalistic theory of mind.
But, promising though it may seem, I'm afraid this line is hopeless, and
for familiar reasons. Design (as opposed to mere order) requires a designer.
Not theologically or metaphysically (pace Paley, Bishop Wilberforce et
al.), but just conceptually. You can't explain intentionality by appealing to
the notion of design because the notion of design presupposes intention-
ality. I do think this is obvious and dreary and that Dennett should give
up trying to swim upstream against it (especially since, as we're about to
see, there's a different route to what he wants that will probably suit his
purpose just as well).
Anyhow, here's the (familiar) argument. Patently, not every effect that
a process has is ipso facto an effect that it designs; short of theology,
at least some effects of every process are merely adventitious. This must
hold of the process of natural selection inter alia. So, in evolutionary
theory as elsewhere, if you wish to deploy the idiom of posed problems
and designed solutions, you must say something about what designing
requires over and above mere causing. Lacking this distinction, everything a
process causes is (vacuously) one of its designed effects, and every one of
its effects is (vacuously) the solution to the problem of causing one of
those.
To be sure, if solutions aren't distinguished &om mere effects, it does
comes out-as Dennett would want it to-that the giraffe's long neck
solved the problem of reaching to the top of things, and did so under
precisely the ecological conditions that giraffes evolved in. But equally,
and in exactly the same sense, I solve the problem of how to make a Jerry
Fodor under the genetic and environmental conditions that obtained in
making me; and the Rockies solve the problem of how to make mountains
just like the Rockies out of just the materials that the Rockies are made of
and under just the conditions of upthrust and erosion in which they
formed; and the Pacific Ocean solves the problem of how to make a hole
of just that size and just that shape that is filled with just that much salt
water; and the tree in my garden solves the problem of how to cast a
shadow just that long at just this time of the day.
5
This is, however, no
metaphysical breakthrough, it's just a rather pointless way of talking; nei-
ther I nor the Pacific get any kudos for being solutions in this attenuated
sense. That's because problems are like headaches; they don't Aoat &ee
of people's having them. The Pacific and I didn't really solve anything
because nobody had the problems that we would have been solutions to.
('Who would want thoser' people always ask.)
178 Chapter 15
Serious talk about problems and solutions requires a serious account of
the difference between designing and merely causing. Notice, moreover,
that if your goal is a reductive theory of intentionality, then your account
of this difference cannot itself invoke intentional idiom in any essential
way. This really does make things hard for Dennett. In the usual case, we
distinguish designing from mere causing by reference to the effects that
the designer did or didn't intend. For example: The flowers Sam gave Mary
made her wheeze and did not please her. They were, nonetheless, a failed
solution to the please-Mary problem, not a successful solution to the
wheeze-Mary problem. That's because Sam intended that receiving the
flowers should please her and did not intend that they should excite her
asthma. Suppose, by contrast, that Mary merely came across the flowers,
and that they both pleased her and made her w h ~ z e . Then the flowers
didn't solve, or fail to solve, anything; they just had whatever effects they
did. I think the intuitions here are about as clear as intuitions can be. It
certainly looks as though the concept of design presupposes, and hence
cannot be invoked to explain, the accessibility of intentional idiom.
If you found a watch on a desert island, you'd have a couple of options.
You could argue that since it was clearly designed, there has to have been
a designer; or you could argue that since there was certainly no designer,
the watch can't have been designed. What is not, however, available is the
course that Dennett appears to be embarked upon: there was no designer,
but the watch was designed all the same. That just makes no sense.
There may be a way out of this somewhere in Dennett's present text or
elsewhere in his writings; but if there is I honestly can't And it. I think
perhaps Dennett is led to underestimate the magnitude of this problem
because he allows himself such free use of the metaphor of Mother Nature
as the designer from whose planning the intentionality of selection flows.
But surely none of that talk can be meant to be literal; reduction has to
stop somewhere, but it can't conceivably stop there. "Jerry Fodor may
joke about the preposterous idea of our being Mother Nature's artifacts,
but the laughter rings hollow; the only alternative views posit one sky-
hook or another" (427). But Dennett doesn't seriously think that having
a scientific world view requires believing that, does he? The story about
Mother Nature is, after all, a fairy tale. There isn't any Mother Nature,
and no intentional agent literally planned me or designed the Rockies or
saw to the tree in the garden. Could Dennett really have lost sight of
that?6
You can't get natural teleology by postulating designerless designs.
Still, both for the sake of the argument and because it may be true, I'm
prepared to grant that a different approach to metaphysical reduction
really might yield a certain kind of natural teleology. Consider a well-
worn example (e.g. Hempel, 1965). Hearts pump blood and they make the
Deconstructing Dennett's Darwin 179
familiar sort of noise, and this correlation is empirically reliable. Nonethe-
less, intuition strongly suggests that hearts were selected for being blood
pumps, not for being noise makers. That is, in such cases intuition reads
"selected for" as intensiona/:
7
The inference from "xs were selected for F"
and "(reliably) F iff G" to "xs were selected for G" is not valid. What, if
not appeals to the intentions of a designer, underwrites such intuitions?
Maybe appeals to counterfactuals do. Maybe, in the present case, the
counterfactual that legitimizes the inference is that hearts would contribute
to fitness (hence would be selected) in (nearly?) worlds where they pump
blood but don't make noise; but they wouldn't contribute to fitness (hence
wouldn't be selected) in (ditto) worlds where they make noise but don't
pump blood. At least this way of getting intensionality into the picture
doesn't invoke the operation of intentional systems (minds);
8
to that
extent it might be of use to a reductivist metaphysical program.
This kind of approach to natural teleology has well-known problems, to
be sure. For one thing, it's not entirely clear that the required counter-
factuals are true. Teleological explanations generally offer sufficient but
not necessary conditions for selection. Maybe if hearts didn't pump, they
would be selected for something else. (Hempel, op. cit.; Dennett recog-
nizes this sort of point in a slightly different context. Seep. 259-260).
For another thing, maybe there just aren't any (nearby?) worlds of the
kind that the counterfactual contemplates. It's arguably a law--hence
nomologically necessary-that hearts pump blood iff they make noise;
and who knows what to make of counterfactuals whose antecedents are
necessarily false. Not only: "who knows what their truth conditions, if any,
are7" but also "who knows what, if any, roles they can play in empirical,
for example, biological, theorizingr'
Digression: Ruth Millikan (1984) has been pushing a line rather different
from Dennett's (if I read her right; which, however, Dennett says I hardly
ever do) according to which what a trait is selected for are those of its
effects that (completely) explain why a creature has it. Correspondingly,
the intensionality of "selected for" derives from the intensionality of
"(completely) explains." (Presumably from :rs being F [completely] explains
y's being G and :rs are F iff :rs are H, it does not follow that :rs being H [com-
pletely] explains y's being G.)
I am pretty comprehensively unmoved by this. Though it's true that
mere extensional equivalence does not license the substitution of "F" for
"G" in F explains that ... , it looks like (nomologically) necessary equiva-
lence does. (See fn. 4.) In particular, if "F iff G" is a law, then for every
explanation according to which trait T is selected for because it causes F,
there will be a corresponding, equally complete, equally warranted expla-
. nation according to which T is selected for because it causes G in a world
where, reliably, Gs are Fs.
180 Chapter IS
The standard philosophical illustration goes like this: Is what explains
the fitness of the frog's feeding reflex that it causes snaps at Aies7 Or is
what explains its fitness that it causes snaps at ambient black dots in cir-
cumstances where it is nomologically reliable that the ambient black dots
are Aies7 Unless there is something to choose between these explanations,
there will be nothing to choose between the corresponding teleological
claims about what the behavioral repertoires of frogs are for. Notice
that, though this example concerns the selection of an intentional trait,
the problem it invokes is entirely general: Appeals to being F can explain
nothing that isn't equally well explained by appeals to being G in a world
where it's necessary that Gs are F.
9
In effect, contexts of explanation are
transparent to the substitution of (e.g., nomologically) necessary equiv-
alents. (For a less sketchy discussion of this group of issues, see Fodor,
1990 and papers in Loewer and Rey, 1991.)
As remarked above, in the present context the main issue is whether a
naturalistic reduction can license intensional teleological idiom. Here, then,
is what's to choose between Millikan's and Dennett's account of selection
for. Dennett can distinguish between selection for nomologically equiv-
alent F and G. That's because he helps himself to "designed for"; and a
thing can be designed for F-ing and not for G-ing even if it's a law that
only Gs F; designing is more intensional than explaining. The down side,
for Dennett, is that there is no design without a designer; so, willy nilly,
he finds himself in bed with Mother Nature. Millikan tries rather harder
not to cheat; explains is (arguably) a relation between propositions, so
(maybe) it doesn't tacitly invoke an agent.
10
The down side, for Millikan,
is that explanation isn't intensional enough; unlike the intensionality of the
mental (and, in particular, unlike the intensionality of design), explanation
doesn't distinguish necessary equivalents. And so neither, of course, does
a notion of selection for that "explains" is used to explain.
Much of a not muchness, if you ask me.
Well, so much about several approaches to natural teleology, and the
qualms that they inspire. I'm sort of inclined to doubt that there is any
natural teleology; I sort of think that the only goals there are the ones
that minds entertain. But never mind. For the sake of the argument, I
hereby grant that there is such a thing as selection for after all, that it can
be naturalistically grounded, and that selection for F doesn't imply selec-
tion for G even if "F iff G" is reliable. The argument will now be that,
even so, the intensionality. of "selection for" doesn't help with the meta-
physics of the intentionality of minds.
Deconstruction
I'm proposing to deconstruct what I take to be Dennett's metaphysics of
intentionality. Now, as everybody knows who's in this line of work, the
Deconstructing Dennett's Darwin 181
way you set about a bit of deconstructing is: You winkle out of the text
some aporia {that means, roughly, glitch, or screw-up) that reveals its deep,
unconscious yen for self-refutation. (Texts are what have yens these days,
the author being dead.) It is in this fashionably postmodem spirit that I
call your attention to what strikes me as distinctly a peculiarity in Den-
nett's views.
On the one hand, it's of the essence of Dennett's program to argue that
there is no principled difference between the intentionality of natural selec-
tion and the intentionality of mind. A principled difference would need a
skyhook to get over it, whereas Dennett's primary effort is to overcome
the "resistance philosophers have shown to evolutionary theories of
meaning, theories that purport to discern that the meaning of words, and
all the mental states that somehow lie behind them, is grounded ulti-
mately in the rich earth of biological function" (402). Or, as he puts it in
a summary passage: "Real meaning, the sort of meaning our words and
ideas have, is itself an emergent product of originally meaningless pro-
cesses-the algorithmic processes that have created the entire biosphere"
(402).
It is because he cares so much about the continuity of the intentionality
of "words and ideas" with the intentionality of selection that Dennett is
so emphatic in denying that the distinction between "derived" and "origi-
nal" intentionality is principled. Since " ... your selfish genes can be seen
to be the original source of your intentionality-and hence of every mean-
ing you can ever contemplate or conjure up," the soi-disant "original"
intentionality of your thoughts is every bit as derived as the patently
parasitic intentionality of thermostats and street signs. Thermostats and
street signs derive their intentionality from us; we derive our intention-
ality from Mother Nature (specifically, from the reproductive ambitions of
our selfish genes). Only the evolutionary process is originally intentional
sans phrase.
That is one part of Dennett's story. But there is another part, equally
strongly insisted on, that might well strike one as distinctly in tension
with the sorts of passages just cited. For, it turns out that " ... Mother
Nature (natural selection) can be viewed as having intentions, [only] in
the limited sense of having retrospectively endorsed features for one reason or
another ... " (462) and ''The Local Rule is fundamental to Darwinism; it is
equivalent to the requirement that there cannot be any intelligent (or far-
seeing) foresight in the design process, but only ultimately stupid oppor-
tunistic exploitation of whatever luck ... happens your way" (191). In this
respect, it appears, Mother Nature is not, after all, much like you and me.
"If [for example] you were playing chess under hidden constraints, you
would adjust your strategy accordingly. Knowing that you had secretly
promised not to move your queen diagonally, you would probably forgo
any campaign that put your queen at risk of capture . . . but you have
182 Chapter 15
knowledge of the hidden constraints and foresight. Mother Nature does
not. Mother Nature has no reason to avoid high-risk gambits; she takes
them all, and shrugs when most of them lose" (259). It is, indeed, pre-
cisely because of her lack of foresight that Mother Nature is forever get-
ting trapped in local maxima: "If only those redwoods could get together
and agree on some sensible zoning restrictions and stop competing with
each other for sunlight, they could avoid the trouble of building those
ridiculous and expensive trunks" (255; Dennett is quoting Dennett 1990,
132).
So, on the one hand, 'There may be somewhat nonarbitrary dividing
lines to be drawn between biological, psychological, and cultural manifes-
tations of this structure [decision making by 'satisficing'], but not only
are the structures-and their powers and vulnerabilities-basically the
same, the particular contents of "deliberation" are probably not locked
into any one level in the overall process but can migrate" (504). But, on
the other hand, Mother Nature is a blind watchmaker; and her blindness
consists (inter alia, see below) in this: Her reasoning is never prudential;
she maximizes her utilities only in retrospect; she never adopts (or rejects) a
policy because of outcomes that she has foreseen-whereas, you and I do
that sort of thing all the time. Why doesn't this difference between her and us
count as principled? I call that an aporia (or do I mean an aporium7 Probably
not).
None of this, please note, takes back so much as a word of what I've
agreed to concede. In particular, I am still reading "select for" and the like
as intensional. But though by assumption Mother Nature's selecting for ...
and my intending to ... have their intensionality in common, I'm now
wanting to emphasize a residual difference between them. Mother Nature
never rejects a trait because she can imagine a more desirable alternative,
or ever selects for one because she can't. We do.
Is this difference principled? I'll argue that it is, anyhow, qualitative; it
shows that the intentionality of minds is a different kind of thing from the
(putative) intentionality of selection. First, however, it's important to get
clear on the scope of the problem. What we've got so far is only the
entering wedge.
It's true but a bit misleading to say that Mother Nature can't foresee
the consequences of selection. For, if she can't see forward, she can't see
sideways or backwards either; all she can see is where she is. So, just as
Mother Nature never prefers F to G because it will (likely) bring it about
that H, so she also never prefers F to G because it would bring about that
H if, counterfactually, it were the case that K; and she never prefers F toG
because Fs but not Gs used to be Js (though they're not anymore). In
short, it's not the (likely) future as such that Mother Nature is blind to; it's
the nonactual as such that she can't see.
Deconstructing Dennett's Darwin 183
In fact, even that underestimates the magnitude of her myopia. Just
as the process of selection is blind to whatever is unactual, so too is it
unaffected by any actual thing that is, de facto, causally isolated from the
process of selecting. Hence, for example, one aspect of the impact of
geography upon evolution: Let it be that Fs are predators on Gs; and let it
be that some genetic variation produces a strain of G* s that are invisible
to Fs. Still, being F-invisible won't provide these G* variants with any
selectional advantages over the old fashioned sort of Gs if, as it happens,
all the Fs live in some other part of the forest. Selectional advantage is the
product of, and only of, real causal interactions. Merely possible com-
petitions don't enter into Mother Nature's calculations; not even if they
are possible competitions between actual creatures. None of this, surely, is
in the least contentious; in fact, it's close to tautology: Nothing is ever the
effect of merely possible causes. Nothing is. Not ever.
Mother Nature never prefers any Fs to any Gs except on grounds of
(direct or indirect) causal interactions that the Fs and G actually have with
the selection process. This looks to me like a whopping difference between
Mother Nature and us. Indeed, if you think of intentionality in the Good
Old Brentano Way, it is exactly what makes mental states intentional and
not just intensional, namely, their capacity for having nonactual objects.
11
Here's a way to summarize the point: Mental states are both intensional
and intentional. Or, if you prefer, mental state predicates resist both sub-
stitution of coextensives and quantifying in. So, in the 6rst case, Lois may
fail to realize that Clark is Kent even though (ho hum!) she knows quite
well, thank you, that Clark is Clark. And, in the second case, Lois may
believe that Santa Claus comes down the chimney even though (ho
hum!) there isn't any Santa Claus. Now, I don't really think that the in-
tensionality of selection explains the intensionality-with-an-s of mental
states. That's because, as I remarked above, however its intensionality
is grounded, "selected for" is likely to be transparent to the substitu-
tion of necessarily coextensive terms, whereas "realizes," "thinks about,"
''believes that," and the like certainly are not. But even if selection did
explain why the mental is intensional-with-an-s, it still wouldn't explain
why the mental is intentional-with-a-t. That's because a state that is
intentional-with-a-t can have "Ideal" or "intentionally inexistent" objects,
and its causal role can (typically does) depend on which Ideal object it has.
Whereas, to repeat, only what is actual is visible to Mother Nature. Only
what is actual can affect the course of a creature's selectional history.
This sort of point is occasionally remarked upon in the literature on
natural teleology; see, for example, Allen and Bekoff (1995): ''For a thing
to possess a biological function, at least some (earlier) members of the
class must have successfully performed the function. In cases of psycho-
logical design, the corresponding claim is not true" (61; my emphasis). That is:
184 Chapter IS
unlike the traits that Mother Nature selects (for), the goals to which
intentional activity is directed can be entirely Ideal. Philosophers whose
project is the reduction of intentionality to natural teleology seem not to
have understood how much difference this difference makes.
But, you might reasonably wish to argue, this can't be a problem for
Dennett's kind of reductionism unless it's a problem for everybody's kind of
reductionism. For, you might reasonably wish to continue to argue, if
nothing can be the effect of a merely Ideal cause, then, a fortiori, thoughts,
decisions, and actions can't be the effects of merely Ideal causes. And sim-
ilarly for other intentional goings on. Remember that Brentano thought
that the intentionality of the attitudes shows that naturalism can't be true;
and Quine thinks that naturalism shows that the attitudes can't be inten-
tional; and the Churchlands think (at least &om time to time) that the
intentionality of the attitudes shows that there can't be any attitudes. If, in
short, there's an argument that Mother Nature is blind to merely possible
outcomes, then there must be the same argument that you and I are blind
to merely possible outcomes. So either Dennett wins or everybody
loses.
Now, I do think that's a puzzle; but it's not one that I just made up. In
fact, it's just the old puzzle about intentionality-with-a-t (as distinct &om
the old puzzle about intensionality-with-an-s). How can what is only Ideal
affect cognitive and behavioral processes? How can I think about, long
for, try to find, blah, blah, the gold mountain even though there isn't any
gold mountain? This is a question which, though it arises for mental pro-
cesses, has no counterpart for processes of natural selection since, to repeat,
although "selected for" and the like are maybe sometimes opaque to the
substitution of equivalents (viz., when the equivalence is not-nomologi-
cally or otherwise-necessary), they are patently always transparent to
existential quantification.
Preliminary moral: You cannot solve the problem of meaning by reduc-
ing the intentionality of mind to the intensionality of selection, because
the intensionality of selection fails to exhibit the very property that
makes meaning problematic. Mother Nature has no foresight, so she can't
ever select for a trait that isn't there. But you can mean something that isn't
there; you can do that any time you feel like.
But then, what sort of story should one tell about intentionality-with-
a-t? There's a commonsense approach which, in my view, probably points
in the right direction but which, no doubt, Dennett would think invokes a
skyhook: Merely intentional objects can affect the outcomes of cognitive
processes qua, but only qua, represented. What appear to be the effects of
Ideal objects on causal processes are invariably mediated by representations
of these objects; and representations, unlike their representees, are actual by
stipulation (they're physical tokens of representation types). The fact that
Deconstructing Dennett's Darwin 185
elephants fly can't make anything happen because there isn't any such
fact. But tokens of the symbol type "elephants fly" mean that elephants
fly, and they can make things happen because there are as many actual
ones of those as you like.
That doesn't, of course, solve the problem of intentionality; it merely
replaces it with the unsolved problem of representation (i.e., of "meaning
that"). Now, I don't know whether representation is a skyhook, and neither
does Dennett. Either there is a naturalistic theory of representation-in
which case, it is the solution to the problem of intentionality-or there is
no naturalistic theory of representation, in which case I, for one, will give
it all up, become an eliminativist about the mental, and opt for early
retirement (well, earlyish). But, either way, the present point stands: You
can't reduce intentionality to "selection for" because selection for doesn't
involve representation. (A fortiori, selection for F-ness doesn't involve rep-
resenting anything as F.) That, in a nutshell, is the difference between
what Mother Nature does when she selects for tall petunias, and what
Granny does when she breeds for them.
12
You can, in short, suppose that the whole (neo-)Darwinian story is true;
and you can suppose that "selection for" is intensional; but you will not
thereby have succeeded in supposing any representation into the world.
And, according to common sense (and according to me), it's representa-
tion that you need to explain intentionality.
Three caveats and we're finished. First, I haven't claimed that only
minds can represent. On the contrary, us informational semanticists are
inclined to think that the representational is a (possibly large) superset of
the intentional. So, for example, it's okay with us if genes, or tree-rings, or
the smoke that means fire have (underived) representational content. It's
just that, according to the previous argument, they don't have it qua
selected.
Second, I hope you don't think that I think that the line of argument
I've been pursuing shows that natural selection couldn't have resulted in
intentional processes. Of course it could; or, anyhow, of course it could
for all that I know. The issue about the "source" of intentionality is not
the historical question "what made intentional things," but the metaphysical
question "what makes intentional things intentional." If what produced
intentionality turns out to have been the action of electricity on the
primal soup, it wouldn't follow that intentionality is either a kind of soup
or a kind of electricity.
Third, a point that should be obvious: It's okay for somebody who is a
reductionist or a naturalist about intentionality to not believe that in-
tentionality reduces to adaptation or natural teleology. Dennett, having
appropriated "naturalistic" for his own brand of reduction, seems unable
to contemplate a view that says "yes, intentionality is reducible, but no
186 Chapter 15
it's not reducible to Darwin." Hence such really bizarre misreadings as" ...
Searle and Fodor ... concede the possibility of [a clever robot] but ... dis-
pute its "metaphysical status"; however adroitly it managed its affairs,
they say, its intentionality would not be the real thing" (426).
But I don't say that, and I never did (and, as I read him, Searle doesn't
and didn't either. I may have mentioned that at one point, Dennett chas-
tises me for being unsympathetic in my reading of him and Millikan. Talk
about your pots and your kettles!) What I have said and still say is that,
however clever a robot (or a creature) is at managing its affairs, its inten-
tionality, its having affairs to manage, does not consist in the cleverness
with which it manages them. If it did, intentionality would supervene on
behavior and behaviorism would be true--which it's certainly not.
I also say that, however clever a robot (or a creature) may be, and
however much its cleverness may, in point of historical fact, have been
the cause of its having been selected, its intentionality does not consist in
its having been selected for its cleverness. Dennett really must see that
saying these sorts of things is quite compatible with being as naturalist
and reductionist as you like about minds. "Jerry Fodor may joke about the
preposterous idea of our being Mother Nature's artifacts, but the laughter
rings hollow; the only alternative views posit one skyhook or another." Non-
sense. There are lots of alternatives to adaptationist accounts of intention-
ality. Some are eliminative and some are reductive; some are naturalistic
and some aren't; some are emergentist and there are even one or two
that are panpsychist. What they have in common is only their being
reconciled to a thought that Dennett apparently can't bring himself to
face, a truly Darwinian idea, and one that is, if not precisely dangerous, at
least pretty disconcerting: There isn't any Mother Nature, so it can't be that
we are her children or her artifacts, or that our intentionality derives from
hers.
Dennett thinks that Darwin killed God. In fact, God was dead a century
or so before Darwin turned up. What really happened was that the
Romantics tried to console themselves for God's being dead by anthro-
pomorphizing the natural order. But Darwin made it crystal clear that the
natural order couldn't care less. It wasn't God that Darwin killed, it was
Mother Nature.
Notes
1. Readers who are puzzled by the epigraph should consult p. 17 of Dennett (1995). All
Dennett references are to that volume except as indicated.
2. What's this "grounding''? Search me; but philosophers, including Dennett, often like to
talk that way. For present purposes, theory A "grounds" theory B only if (i) the ontol-
ogy of A reduces the ontology of B, and (ii) A is independently justified.
Deconstructing Dennett's Darwin 187
3. Likewise, the organism's phenotypic constellation of psychological traits and capadties
is Altered and transformed by whatever processes "read" it onto the creature's behavior.
Hence the need for a "performance/competence" distinction within theories at the psy-
chological level.
4. See Sober (1984), whose generally admirable treatment of these issues is, however, mis-
taken at a crudal point. Sober claims that selection for-but not selection-is intensional,
and also that what gets selected is objects, whereas what gets selected for is traits (proper-
ties). But, as we're about to see, selection for contexts are trlensiorull (viz., transparent, like
seltction contexts) in respect of the substitutivity of coextensive trait terms when the
coextension is (e.g., nomically) necessary. That, in fact, is one of the main reasons that
seltction for is so unlikely to ground a theory of the intensionality of the mental.
5. Moreover, all these solutions meet Dennett's requirement of being "algorithmic." It is,
however, unsurprising that they do so since, as Dennett rightly says, there aren't "any
limits at all on what may be considered an algorithmic process .... " (59) Step right up
and play; every contestant is a guaranteed winner.
6. It adds to the confusion that Dennett often writes as though ''Mother Nature" is just a
figurative name for natural selection. That is, of course, perfectly okay; but you can't both
talk that way and also hold, as a substantive thesis, that her intentionality explains its.
7. The wary reader will notice that I have switched from intentionality-with-a-t to
intensionality-with-an-s. Quite so; the point of this will appear in due course.
8. Unless the truth makers for counterfactuals are themselves mind-dependent, as, indeed,
philosophers have sometimes supposed.
9. It's unclear how much of this sort of teleological or intensional indeterminacy Millikan is
prepared to live with. Dennett, however, is explidt that he doesn't mind if there is no
fact of the matter about the intentional object of the frog's snap. But he doesn't say
whether he's equally sanguine about every ascription of intentional content to the frog
being indeterminate between reliably coextensive properties. Nor does he say why, on
his view, the same wouldn't hold of intentional ascriptions to us. (Do not reply "that's
because we have language" or I shall break down and commence to gnaw the rug:
Exactly the same indeterminacy would infect the assignment of meanings to reliably
coextensive predicates of, say, English.)
10. I'm being very nice and assuming, concessively, that explanation isn't itself a pragmatic
notion. If what explains what depends on who wants to know what, then Millikan's line
is just as question-begging as Dennett's.
11. This last is by way of cashing footnote 7.
12. Sober (op. dt., p. 202) remarks that "Artifidal selection is a variety of natural selection;
the relation is one of set inclusion, not disjointness." Well, yes and no. It's quite true that,
if you are typing selectors by the effects of their sorting, Granny and Mother Nature are
both just Alters on phenotypic variance. But if you are typing selectors by how they achieve
their effects, the difference could hardly be more striking. In the whole natural order, as
far as anybody knows, only Granny and her kind Alter with. literally, premeditation.
Chapter 16
Is Science Biologically Possible? Comments on Some
Arguments of Patricia Churchland and of Alvin
Plantinga
I hold to a philosophical view that, for want of a better term, I'll call by
one that is usually. taken to be pejorative: Scientism. Scientism claims, on
the one hand, that the goals of scientific inquiry include the discovery of
objective empirical truths; and, on the other hand, that science comes
pretty close to achieving this goal at least from time to time. The molecu-
lar theory of gasses is, I suppose, a plausible example of achieving it in
physics; so is the cell theory in biology; the theory, in geology, that the
earth is very old; and the theory, in astronomy, that the stars are very far
away. Scientism is, to borrow a phrase from Hilary Putnam the scientist's
philosophy of science. It holds that scientists are trying to do pretty much
what they say that they are trying to do; and that, with some frequency,
they succeed.
I'm inclined to think that scientism, so construed, is not just true but
obviously and certainly true; it's something that nobody in the late twen-
tieth century who has a claim to an adequate education and a minimum of
common sense should doubt. In fact, however, scientism is tendentious.
It's under attack, on the left, from a spectrum of relativists and pragma-
tists, and, on the right, from a spectrum of Idealists and a priorists. People
who are otherwise hardly on speaking terms-feminists and fundamen-
talists, for example-are thus often unanimous and vehement in rejecting
scientism. But though the rejection of scientism makes odd bedfellows, it
somehow manages to make them in very substantial numbers. I find it, as
I say, hard to understand why that is so, and I suppose the Enlightenment
must be turning in its grave. Still, over the years I've sort of gotten used
to it.
Recently, however, the discussion has taken a tum that seems to me
bizarre. Scientism is now being attacked by, of all people, Darwinists. Not
all Darwinists, to be sure; or even, I should think, a near majority. Still,
the following rumor is definitely abroad in the philosophical community:
When applied to the evolution of cognition, the theory of natural selec-
tion somehow entails, or at a minimum strongly suggests, that most of our
empirical beliefs aren't true; a fortiori, that most of our empirical scientific
190 Chapter 16
theories aren't true either. So the rumor is that Darwinism-which, after
all, is widely advertised as a paradigm of scientific success (I've heard it
said that Darwinian adaptationism is the best idea that anybody' s ever
had, and that natural selection is the best confirmed theory in science)-
Danoinism, of all things, undermines the scientific enterprise. Talk about
biting the hand that feeds you!
There isn't, even among those who propagate this rumor, much con-
sensus as to what one ought to do about it; indeed, there's currently a
lively philosophical discussion on just that topic. Some say "well, then, so
much the worse for truth." Others say "well, then, so much the worse for
science." Still others say, "well, then, the paradox must be merely appar-
ent; nothing true can refute itself."
In this discussion, I'll take a different line &om any of these. My theme
is that we don't need to worry about what to do if Darwinism threatens
scientism because, in fact, it doesn't. There is, I claim, nothing at all in
evolutionary theory that entails, or suggests, or even gives one grounds
to contemplate denying the commonsense thesis that scientific inquiry
quite generally leads to the discovery of objective empirical truths. I'll try
to convince you of this by looking, in some detail, at the arguments that
are supposed to show the contrary. I hope you won't find the process
tedious; this sort of inquiry can be edifying and, anyhow, it's what philos-
ophers mostly do for a living.
Perhaps I should mention, in passing, that I'm not actually an adapta-
tionist about the mind; I doubt, that is, that Darwinism is true of the
phylogeny of cognition. For reasons I've set out elsewhere (see the two
preceding chapters) I think our kinds of minds are quite likely ''hopeful
monsters," which is why there are so few of them. It would thus be open
to me to take the line that if scientism is jeopardized by adaptationism
about the evolution of the cognitive mind, so much the worse for adapta-
tionism about the evolution of the cognitive mind. But, though I could do
that, on balance I'd rather not. That's because I'm much more confidant
that scientism is true than I am that Darwinism about the phylogenesis of
cognition is false. I want to leave open the possibility that they might
both be true. I'm not, therefore, about to concede that they aren't com-
patible. For present purposes, then, I propose to suppress my qualms and
pretend that I believe that our minds evolved; and that they did so under
selection pressures. The question of interest is what, if anything, this
implies about the likely truth of scientism.
So, what is the argument &om adaptationism about cognition (hence-
forth, for short, the argument &om "Darwinism") to the rejection of
scientism7 Actually, it's an inference that comes in several different pack-
ages. I propose to look at two. Here, to begin with, is a quotation &om
Patricia Churchland (1987): ''There is a fatal tendency to think of the brain
Is Science Biologically Possible? 191
as essentially in the fact-finding business .... Looked at from an evolution-
ary point of view, the principle function of nervous systems is to get the
body parts where they should be in order that the organism may sur-
vive .... Truth, whatever that is, definitely takes the hindmost."
Clearly, the crux of this quote is the claim that "the principle function
of nervous systems is [not the fixation of true beliefs, but] to get the body
parts where they should be in order that the organism may survive." Just
why does Churchland think that "the evolutionary point of view" implies
this? It's not, of course, that there's been some hot new paleobiological
discovery she's just heard about (''Flash: Fossil Nervous System Found
Functioning in China"). Rather, what she apparently has in mind is some
inference from Darwinist general principles. Perhaps it's this: Organisms
get selected for getting their bodies where they should be; so nervous sys-
tems get selected for getting the bodies of organisms to where they
should be; so the function of nervous systems-the function for which
selection designed them-is to get the bodies of organisms to where they
should be. A fortiori, the function of nervous systems isn't "fact-finding."
But if that is the intended inference, it commits a sort of distributive
fallacy; it's like arguing that, since the army is numerous, each of the sol-
diers must be numerous too. Consider the following parody: "Selection
pressures favor reproductive success; the design of the heart was shaped
by selection pressures; so the function of the heart is to mediate repro-
ductive success." Parallel arguments would show that the function of every
organ is to mediate reproductive success; hence that every organ has the
same function.
Poppycock! The function of the heart-the function that its design
reflects-is to circulate the blood. There is no paradox in this, and noth-
ing to affront even the most orthodox Darwinist scruples. Animals that
have good hearts are selected for their reproductive success as compared
to animals that have less good hearts or no hearts at all. But their repro-
ductive success is produced by a biological division of labor between their
organs, and it's the function of the heart relative to this division of labor that
the heart was designed to perform.
Likewise, it's entirely possible that the kind of mental architecture that
maximizes behavioral adaptivity is also one that institutes a division of
labor: Perhaps a cognitive mechanism that is specialized to figure out
what state the world is in interacts with a conative mechanism that is
specialized to figure out how to get what one wants from a world that is
in that state. That's, roughly, the sort of nervous system that you'd expect
if creatures act out of their beliefs, desires, and practical intelligence; and
there's nothing in Darwin to suggest that they do otherwise. All Darwin
adds is that, if that is how a creature works, then, on average, it will work
better, and last longer, and have more progeny in proportion as its beliefs
192 Chapter 16
are true and its desires are appropriate to its ecological condition. Who
would deny it7
Churchland's argument doesn't convince. But maybe there's better on
offer? Here's a line that Alvin Plantinga (forthcoming; all Plantinga refer-
ences are to this manuscript) has recently been pushing:
1
"Our having
evolved and survived makes it likely that our cognitive faculties are reli-
able and our beliefs are for the most part true, only if it would be impos-
sible or unlikely that creatures more or less like us should behave in
fitness-r:nhancing ways but nonetheless holds mostly false beliefs" (5).
But, in fact, it isn't impossible or unlikely that creatures more or less like
us should behave in fitness-enhancing ways but nonetheless hold mostly
false. beliefs. "[T]he simplest way [of seeing how evolution might get away
with selecting for cognitive mechanisms that produce generally false
beliefs] is by thinking of systematic ways in which . . . beliefs could be
false but still adaptive .... [Imagine a creature that] thinks all the plants and
animals in [its] vicinity are witches, and [whose] ways of referring to them
all involve definite descriptions entailing witchhood .... [T]his would be
entirely compatible with [its] beliefs' being adaptive; so it is clear ... that
there would be many ways in which [its] beliefs could be for the most part
false, but adaptive nonetheless." (10).
It looks like the argument that Plantinga has on offer is something like
this:
Argument A: Our minds evolved, so we can assume that our
behavior is mostly adaptive. But our behavior could be mostly
adaptive even if our beliefs were mostly false. So there's no reason
to think that most of our beliefs are true.
At worst, Argument A seems better than Churchland's since it con-
cedes, straight off, that the mind is a belief-making mechanism: Plantinga' s
point is that you can imagine the mind's succeeding as a belief-making
mechanism even if the beliefs that it makes aren't true. Still, I don't think
that Argument A could be exactly what Plantinga has in mind; not, at
least, if what he has in mind is an argument against scientism (or against
what he calls "naturalism," which, at least for present purposes, we may
take to be much the same thing). But I do think that there's quite an inter-
esting argument in the immediate vicinity of Argument A; one that has
every right to be taken seriously. We'll come back and fix this presently.
First, however, I want to grumble about a bit about a premise.
It's important to Planting a that it not be in doubt that a system of
mostly false beliefs could be adaptive. If there were anything wrong with
the notion of a system of mostly false but adaptive beliefs, Plantinga' s
Darwinian argument against scientism couldn't get off the ground. Plan-
tinga' s way of assuring us that there's nothing wrong with this notion is
Is Science Biologically Possible? 193
to sketch a general procedure for constructing such a system: Start with
our beliefs, but replace all the referring expressions by complex demon-
stratives (or definite descriptions) with false presuppositions. We think
that tree is VerY big; they think that witchtree is very big; we think the cat on
the mat is asleep; they think the witchcat on the witchmat is asleep. And so
on. What you end up with is a creature whose behavior is plausibly about
as adaptive ours, but most of whose beliefs are false.
Now I do think that goes a little fast. After all, one needs to explain
why the behavior of a creature thus epistemically situated would be
largely successful; and the intuition, surely, is that, on any way of count-
ing them that's defensible, "most of the beliefs" that mediate its behav-
ioral successes tum out to be true after all. Let it be that that appletree
witch is blooming is false, or lacks a truth value, in virtue of its presuppos-
ing that that appletree is a witch. Still, much of what a creature believes
in virtue of which it believes that that appletree witch is blooming (and in
virtue of which the thought that that appletree witch is blooming leads to
behavioral successes) are perfectly straightforwardly true. For example:
that's an appletree; that's blooming; that's there; something is blooming; some-
thing blooming is there, and so on indefinitely. The point is trivial enough:
If a creature believes that appletree witch is blooming, then it presumably
believes that that's an appletree and that that's a witch and that that's
blooming. And two of these are true beliefs that the creature shares with
us, and that enter into the explanation of its behavioral successes vis-a-vis
blooming appletrees in much the same way that the corresponding beliefs
of ours enter into the explanation of our behavioral success vis-a-vis
blooming appletrees. Notice that this argument doesn't depend on assum-
ing (as, in fact, I don't) that belief is dosed under implication, or even
under "immediate" implication. All it needs is the truism that the creature
has whatever beliefs the explanations of its behavioral successes require
attributing to it. This, surely, is common ground whether or not the beliefs
that such explanations appeal to are supposed to be true.
The philosopher Berkeley seems to have believed that tables and chairs
are logically homogeneous with afterimages. I assume that he was wrong
to believe this. But intuition strongly suggests that the truth of "most of"
his beliefs survived his metaphysical aberrations, and that it was these
residual true beliefs that explain his behavioral successes. He may, for
example, have chosen to sit in this chair rather than that one because he
believed that this afterimage chair is more comfortable than that after-
image chair. And that may have been the right choice in the circumstance.
If so, we should credit Berkeley's success to his true belief that this chair is
comfortable, and not to his false belief that this chair is an afterimage.
The long and short is: There's admittedly no agreed-on way of count-
ing beliefs; but it's not dear that the beliefs of the creature Plantinga asks
194 Chapter 16
us to imagine are mostly false. Or, indeed, that its epistemic situation
actually differs very much &om our own. So, suppose that the forces of
natural selection can't distinguish that creature &om us; suppose, that is to
say, that its aggregate beliefs are about as adaptive as ours. Even so, so far
we don't have a crystal dear case of the kind that Plantinga needs: one
where selection might favor a creature whose beliefs are, in some radical
sense, "mostly false."
If I'm right about this, it leaves us at a bit of an impasse. I'm pretty sure
that Plantinga' s way of constructing a system of beliefs that are dearly
mostly adaptive but mostly false doesn't work. On the other hand, I
haven't an argument for what I strongly suspect, that there is no way of
constructing such a system. And I don't want the discussion to break
down here, because I think the following question is worth considering: If
it is possible to construct a system of adaptive but mostly false beliefs, is there
then an argument from Darwinism to the rejection of scientism?
So here's what I propose. Let's just assume that there is a "charmed
circle" of nonstandard cognitive mechanisms such that:
every cognitive mechanism in the charmed circle produces
systems of mostly false beliefs, and
most creatures that act on such systems of false beliefs (most
"false believers" as I'll say) will do at least as well as most creatures
that ad on the corresponding true beliefs (i.e., at least as well as
most corresponding "true believers").
It is very nice of me to concede all this since I really am much inclined
to doubt that it's true.
We need to do one more preliminary thing before we can get to the
serious stuff. As we've seen, Plantinga writes that "Our having evolved
and survived makes it likely that our. cognitive faculties are reliable and
our beliefs are for the most part true, only if it would be impossible or
unlikely that creatures more or less like us should behave in fitness-
enhancing ways but nonetheless hold mostly false beliefs" (5).
Now, that's okay as it's stated because, as it's stated, it only purports to
give a necessary condition for
(1) Our having evolved makes it likely that our cognitive faculties
are reliable.
But what Plantinga actually needs, if he's to have an argument against
scientism, is something like a sufficient condition for
(2) Our having evolved makes it likely that our cognitive faculties
are unreliable.
Is Science Biologically Possible? 195
Short of arguing for (2), the most Plantinga' s got is that Darwinism is
compatible with scientism' s being false; which I suppose to be true but not
interesting. That the cat is on the mat is also compatible with scientism' s
being false. So what? The cat' s being on the mat isn't therefore a reason to
think that scientism is false. Churchland appears to make a similar mistake:
If evolution selected our minds for moving our bodies, then perhaps Dar-
winism gives us no reason to suppose that our beliefs are mostly true. But
that's quite different from Darwinism giving us some reason to suppose
that our beliefs are mostly not true, and it's this latter than the Darwinian
argument against scientism requires.
In short, what Plantinga really needs (and what I haven't any doubt is
what he really intends) is argument A . Argument A is just like Argu-
ment A except that the conclusion is stronger.
Argument A*: Our minds evolved, so we can assume that our
behavior is mostly adaptive. But our behavior could be mostly
adaptiye even if most of our beliefs were mostly false. So there is
reason to think that most of our beliefs are false.
Now, if I were in a bad mood, or if I were generally an unobliging chap,
I could, with some justice, harp on the difference between arguments A
and A . Consider:
Argument B: I am standing here giving a lecture and I am wearing
my gray socks. But there are other colors (green, say) such that,
quite likely, someone a lot like me might be standing here giving a
lecture if he were wearing socks of any of those colors. So that I am
wearing gray socks is not a reason to believe that I am standing here
giving a lecture.
Argument B seems alright; that I am wearing gray socks isn't a reason
to believe that I am standing here giving a lecture. It's just that, as a matter
of fact, both happen to be true. But, notice, argument B* doesn't seem
right at all.
Argument B*: I am standing here giving a lecture and I am wearing
gray socks. But there are other colors of socks (green, say) such that,
quite likely, someone a lot like me might be standing here giving a
lecture if he were wearing socks of any of those colors. So that I am
wearing gray socks is a reason to believe that I am not standing here
giving a lecture.
Poppycock once again. But, as far as I can see, the difference between
Band B* is just like the difference between A and A*, and it's important
not to equivocate between A and A for just the same reason that it is
196 Chapter 16
important not to equivocate between B and B*. I'm a little worried that
Plantinga' s Darwinian argument against scientism trades on this sort of
equivocation; I haven't, anyhow, found a way to make it run that clearly
doesn't.
So, as I say, if I were in a bad mood or generally unobliging I could
make a fuss of this. But I'm not going to because, as it turns out, my main
line of objection to Plantinga is good against both A and A. Nothing that
follows, therefore, will depend on the difference between them (except at
the very end).
Recapitulation: I've agreed to assume that the charmed circle is non-
empty. By this assumption, it's likely that there could be false believers
that would, in principle, be just as successful as true believers are. What I
need to show is that, pace Plantinga, this assumption doesn't even come
close to underwriting either A or A .
Well, it doesn't. Here's why: What I've just conceded is, in effect, that
what evolution "really wants" is not true believers, but either true believers
or just any old believers whose cognitive mechanisms are in the charmed
circle. But the question nonetheless arises, how, in our case, did evolution go
about getting what it wanted? Darwinism is, remember, a historical thesis
about us. If it's to jeopardize scientism, it's got to underwrite the claim
that our minds actually evolved under selection that favored us because
our false beliefs were generally successful. That there are, in principle,
systems of generally successful false beliefs isn't enough to show this;
nor is it enough that there quite possibly could be creatures a lot like us
whose minds evolve under selection for false beliefs and whose behaviors
are generally successful. What's required, to repeat, is a plausible historical
scenario according to which our minds were selected because we had lots
of behaviorally successful false beliefs. Well, is there such a scenario? The
answer depends on the status of the following claim:
There is some property P of our beliefs such that:
i. In the past, our acting on false beliefs that have P actually has
been (at least) as successful as our acting on true beliefs was. (Other-
wise selection would have had a reason for preferring true beliefs to
false beliefs with P.)
ii. It's reliable (counterfactual supporting) that false beliefs with
P are at least as behaviorally successful as true beliefs. (That is, at
least some cognitive mechanisms that produce P beliefs are in the
charmed circle.)
iii. Assuming our cognitive system is still doing what it evolved to
do-that is, still performing the function for which it was selected-
our current behavioral successes are still largely predicated on false
beliefs that have P.
Is Science Biologically Possible? 197
The reason there has to be such a property is simply that, according to
the present story, evolution selected us because of something about lots of
our false beliefs that reliably lead to our behavioral success. So the present
story couldn't be true unless there was something about lots of our false
beliefs that reliably lead to our behavioral successes.
Well, was there? I doubt it, and (short of philosophical skepticism),
nobody I've heard of actually believes it. What we all believe is that when
actions out of false beliefs are successful, that's generally a lucky accident;
and, correspondingly, that a policy of acting on false beliefs, even when it
works in the short run, generally gets you into trouble sooner or later.
Indeed-and this is the point I want to stress-the commonsense view is
that we have, even as we stand here now, lots of solid inductive evidence
that our false beliefs pretty generally don't have property P. The induc-
tive evidence is that, though some of our actions on false beliefs have
succeeded &om time to time, pretty generally our actions on false beliefs
have failed; and they've failed precisely because they were actions on false
beliefs. Conversely, in what must surely be overwhelmingly the standard
case, if the explanation of a behavioral success invokes a creature's beliefs
and desires at all, it does so as follows: The creature acted in such and
such a way because it wanted to bring it about that so and so; and
because it believed, truly, that acting that way would bring it about that
so and so. That much about intentional explanation has been clear since
Aristotle. Perhaps, when you're doing philosophy, you're prepared to
doubt it; but you don't, surely, doubt it when you're being serious.
This commonsense view of the relation between true belief and suc-
cessful behavior may be wrong, of course, but I want to stress that you
can't invoke the assumption that it's wrong on behalf of the putative
Darwinian argument against scientism. For, that we are probably false
believers is supposed to be the conclusion of the Darwinian argument, not
its premise. It is very easy when one is doing epistemology-especially
when one is doing skeptical epistemology-to find oneself arguing back-
wards. I want to pause to rub this point in a bit, because I think it may
really be at the bottom of a lot of the trouble.
You clearly could get the required antiscientistic Darwinian conclusion
&om the following argument assuming that its premises are true.
Argument C: 1. Many (/most) of our behavioral successes probably
are (/probably have been) predicated on false beliefs that have P.
2. So selection has had inductive grounds for favoring false beliefs
that have P. 3. So we've probably evolved to have false beliefs that
have P.
But, surely, Argument C is question begging in the present context. As
it stands it's no better than the first premise, and on the face of it, the first
198 Chapter 16
premise is not plausible. What's needed to make the argument convincing
is some independent reason for thinking that many of our behavioral suc-
cesses actually have been predicated on false beliefs that have P. And, on
pain of circularity, it better be a reason that doesn't itself assume that
we've probably evolved to have false beliefs. Short of philosophical skep-
ticism, I know of no independent reason to believe premise 1 and, as
remarked above, all the inductive evidence I've heard of is to the contrary.
I hope the main line of my argument against Plantinga is now getting
clear. It goes in three steps.
Step 1: Believing that there is no such property as Pis not at all the
same as believing that the charmed circle is empty, or believing that it's
likely that the charmed circle is empty. I, for example, believe that, in
point of historical fad, those of my successful actions that have been pre-
dicated on false beliefs were mostly flukes; a fortiori, I believe that they
weren't predicated on false beliefs that had P. But I also believe there
could be cognitive systems of the sort that membership in the charmed
circle requires. All that needs is some property such that if my false beliefs
had had it, my actions would have been reliably successful. It's quite con-
sistent of me to believe that there very likely are properties such that if
my false beliefs had had them they would have been reliably successful,
but to doubt that the false beliefs on which my successful actions have
actually been predicated actually have had any of these properties. And
likewise for the false beliefs on which the successful actions of my grand-
mother were predicated; and likewise for those of her grandmother; and
likewise back to the protoplasmic slime.
Step 2: For it to be true that I was selected because in the past my false
beliefs had P, it would at least have to be true that a lot of my past
behavioral successes have been predicated on false beliefs tout court. But,
in fad, I have no reason at all to believe that lots of my past behavioral
success have been predicated on false beliefs, and I have vast lots of
inductive evidence to the contrary. For example, this morning I managed
to get my teeth brushed; a small behavioral success, to be sure, but mine
own. Such as it was, I'm quite certain that it was prompted and guided
by a host of true beliefs including, inter alia: true beliefs about my teeth
needing a brush; true beliefs about the spatiotemporal location of my
toothbrush; true beliefs about the spatiotemporal location of my teeth;
true beliefs about the spatiotemporallocation of my limbs with respect to
my tooth brush and my teeth, and so on. Certainly, short of philosophical
skepticism, I can think of no reason in the world to deny any of this.
I pause to emphasize the caveat. I said that I can think of no reason
short of philosophical skepticism for doubting the commonsense view that
most of one's successful actions are and have been actions out of true
beliefs. This way of putting it is not question begging in the present con-
Is Science Biologically Possible? 199
text. For, I am not trying to refute philosophical skepticism, nor do I
doubt that Darwinism together with philosophical skepticism would argue
that we probably aren't true believers. That must be so since philosophical
skepticism, all by itself, would argue that we probably aren't true believers
if only it were true. None of that is pertinent, however, in the present
context; the question before the house is whether Darwinism undermines
scientism from, as it were, the inside; namely, without supplementary
philosophical commitments.
Step 3: If the mere nonemptiness of the charmed circle proves nothing,
and if the available inductive evidence is that our successful actions are
quite generally predicated on true beliefs, then all that's left for the Dar-
winian argument against scientism is to suppose that things have changed
a lot since our minds got selected: Whatever the situation is now, it used
to be that our successful behavior was generally predicated (not on true
beliefs but) on false beliefs that have P. That's how it was back when
selection molded cognition.
But why on earth should anyone believe this? And, even if anyone does
believe it, why does it matter in the present context? That minds used to
achieve their successes by causing us to act on false beliefs is no challenge
to scientism as long as they don't do it that way any more. The relevant con-
sideration is one that Gould and Lewontin stressed in their famous paper
about spandrels: "(there are cases where one finds] adaptation and selec-
tion but the adaptation is a secondary utilization of parts present for rea-
sons of architecture, development, or history. If blushing turns out to be
an adaptation affected by sexual selection in humans, it will not help us
to understand why blood is red" (1979, 159). Correspondingly, even if it
turns out that nervous systems were originally selected for making adap-
tive false beliefs-a hypothesis which I don't believe for a minute, and
which we've thus far been offered no reason to grant-that still wouldn't
ground a Darwinian argument against scientism.
Now we're ready to recap. The polemical situation, according to the
opposition, is that a Darwinist has no reason for supposing that evo-
lution picks true believers rather than false-believers-with-cognitive-
mechanisms-in-the-charmed-circle. This is because, by assumption, the
behavior of the latter would be just as adaptive as the behavior of the
former. But, I claim, that's not good enough. It can't be plausible that we
were selected for the behavioral success of our false beliefs unless it is
independently plausible (or, anyhow, independently not implausible) that a
lot of our behavioral successes have in fact been actions out of false
beliefs. But, to repeat, the inductively plausible commonsense view is that,
however we may have evolved, our behavioral successes are generally
predicated on true beliefs; and that it's generally the truth of the beliefs on
which they are predicated that explains the successes of our behaviors;
200 Chapter 16
and that, in cases when we have managed to act successfully out of false
beliefs, that's mostly been blind luck. As far as I can tell, there is no reason
on God's green earth why a proper Darwinism should reject this com-
monsense view.
I propose to conclude with a caveat. It's been one of my complaints
against Churchland and Plantinga that they seem to be inclined to argue
from "Darwinism doesn't imply that most of our beliefs are true," to "it's
therefore likely that most of our beliefs are false"; an argument which is,
as I remarked, invalid on the face of it. But I also want to warn against the
converse fallacy, that of arguing from "Darwinism doesn't imply that
most of our beliefs are false" to "it's therefore likely that most of our
beliefs are true."
A fair lot of philosophers have been tempted by the idea that, far from
refuting scientism, the theory of evolution actually underwrites it. Thus
Dan Dennett writes (1978a, 17), "[T]he capacity to believe would have no
survival value unless it were a capacity to believe truths." What he's
endorsing is not the casual truism that any creature that is able to believe
falsely that P is likewise able to believe truly that not-P; it's the much
stronger claim that the Darwinian survival of a thinking creature ipso
facto implies, or and how makes plausible, the truth of most of what that
creature believes.
But that simply isn't so, and the reason is much the one that Plantinga
suggests. That a creature evolved to have some organ is, if the Darwinian
story is true, something like a sufficient argument that having that organ
was adaptive for the creature. But it is not part of the Darwinian story,
nor is it independently the case, that if a creature failed to evolve to have
some organ, that's anything like a sufficient argument that having it
wouldn't have been adaptive for the creature. Darwinism deals in sufficient
conditions for adaptivity, not in necessary conditions for adaptivity.
2
The
point this observation turns on is, in fact, perfectly general: Darwinism
isn't about which traits of organisms would have been adaptive if a creature
had had them. Darwinism is about what effects the actual possession of a
trait has as a function of its adapativity: roughly, the frequency with which
a trait is represented in a population is proportional to its adaptivity, all
else equal.
To put much the same point the other way round: What Darwinism
requires as (roughly) sufficient for the evolution of a trait is both that
the trait would be adaptive if the organism were to have it; and that, in
point of historical fact, there are some (rudimentary, approximate, proto-)
instances of the trait actually on the ground for selection to operate upon.
I've supposed that what was on the ground when our minds evolved
were primitive cognitive mechanisms that produced, by and large, adap-
tive true beliefs; and that selection favored, among such cognitive mecha-
Is Science Biologically Possible? 201
nisms, the ones that produced adaptive true beliefs efficiently and reliably
and about a wide range of ecologically salient states of affairs. The reason
I've supposed this is that, pretty clearly, it's mostly true beliefs that even-
tuate in adaptive behavior now. But, to repeat, Darwin would have been
just as happy if what had been on offer for selection to choose from when
we evolved was cognitive mechanisms which produce, by and large,
adaptive false beliefs; in that case, the Darwinian prediction would have
been that it's mostly false beliefs that eventuate in adaptive behavior
now. It is, just as Plantinga and Churchland both say, the adaptivity of the
beliefs, and not their truth per se, that evolution cares about. Since, for all
we know, evolution would have chosen false believers if it had been
given the chance, the fact that evolution chose us isn't, in and of itself, a
reason for thinking that we're true believers. There is, as far as I can tell,
no Darwinian reason for thinking that we're true believers. Or that we
aren't.
The problems of philosophy are very hard, so we keep hoping that
somebody else will take them off our hands. Darwin, Heaven help him,
has been for years among our favorite candidates. In the nineteenth cen-
tury, we thought that maybe he would do our ethics for us. These days,
it's the burden of grounding our epistemology that we want him to carry.
According to temperament, we'd either like him to show that most of our
beliefs have to be true or that most of them have not to be. Either way,
we want Darwin to demarcate, once and for all, the necessary limits of
human fallibility. He's done so much for biologists; why can't he do this
one little thing for philosophers?
But Darwin isn't in the epistemology business, and evolution doesn't
care whether most of our beliefs are true. Evolution is neutral as to
whether most of our beliefs are true. Like Rhett Butler in the movie, it just
doesn't give a damn.
So, then, why do I believe in scientism? Or, to put the question slightly
differently: If the theory of natural selection doesn't imply that our empiri-
cal beliefs are generally pretty close to being true, then why, in particular,
do I believe that the lunar theory of the tides, and the tectonic theory of
the arrangement of the continents, are pretty close to being true? Put
that way, the question sort of answers itself. I believe these theories on
account of my (casual, to be sure) acquaintance with the evidence that has
been alleged for them. And, of course, the evidence that's been alleged for
the lunar theory of the tides isn't that human cognition evolved; it's that,
for example, the lunar theory correctly predicts that (all else equal) the
tides are highest when the moon is full. Likewise, the evidence alleged for
the tectonic theory of the arrangement of the continents isn't that human
cognition evolved; it's that, for example, the geology of Western Africa
matches the geology of Eastern South America in ways that are otherwise
202 Chapter 16
quite inexplicable. This is just what you'd expect; and it's just what the
scientists themselves take for granted. What on earth does whether or not
our minds evolved have to do with whether or not it's the moon that
pulls the tides around?
Maybe, though I very much doubt it, the whole scientific enterprise
somehow requires, or presupposes, an epistemological foundation to rest
on. And if philosophers can't get it one a priori by an exercise of Carte-
sian intuition, maybe they can get it one some other way. Well, maybe.
But Darwin isn't going to help. On the contrary, since the theory of evo-
lution is itself a part of the scientific enterprise, whatever epistemological
boat we're in, Darwin's in there with us. My scientism is not in the least
disturbed by this. Quite the opposite, it's perfectly delighted to have him
aboard.
Acknowledgment
I'm grateful to Professor Plantinga for some email exchanges about this
material, and for allowing me to read Plantinga (forthcoming) in a manu-
script version. Plantinga doesn't think I've got to the bottom of the argu-
ments he gives there. No doubt he's right.
Notes
1. There is, in fad, more to Plantinga's argument than what I will discuss here. Plantinga
thinks (as I do not) that there is some serious chance it will turn out that we don't act out
of the content of our mental states Ill all, and a fortiori that the truth of our beliefs can't be
what explains our behavioral successes. In any event, I'll consider only the aspects of
Plantinga' s polemic that turn on Darwinst considerations.
2. This is quite a general property of teleological explanations, as, by the way, Carl Hempel
pointed out some while ago (HempeL 1965).
Chapter 17
Review of Steven Pinker's How the Mind Works and
Henry Plotkin's Evolution in Mind
It belongs to the millennial mood to want to sum things up, see where we
have gotten, and point the direction where further progress lies. Cogni-
tive science has not been spared this impulse, so here are two books pur-
porting to limn the state of the art. They differ a bit in their intended
audience; Plotkin's is more or less a text, while Pinker hopes for a general
readership. Pinker covers much more ground, but he takes an ungainly
600 pages to do it, compared to Plotkin's svelte 275. Both authors are
unusually good at exposition, Pinker exceptionally so &om time to time.
And their general sense of what's going on and of what should come next
are remarkably similar considering that they are writing about a field that
is notoriously &actious. Taken severally or together, they present what is
probably the best statement you can find in print of a very important
contemporary view of mental structure and process.
But how much of this view is true7 To begin with, Pinker and Plotkin
are reporting a minority consensus. Most cognitive scientists still work in
a tradition of empiricism and associationism, whose main tenets haven't
changed much since Locke and Hume. The human mind is a blank slate at
birth. Experience writes on the slate, and association extracts and extrap-
olates whatever trends there are in the record that experience leaves. The
structure of the mind is thus an image, made a posteriori, of the statistical
regularities in the world in which it finds itself. I would guess that quite a
substantial majority of cognitive scientists believe something of this sort;
so deeply, indeed, that many hardly notice that they believe it.
Pinker and Plotkin, by contrast, epitomize a rationalist revival that
started about forty years ago with Chomsky's work on the syntax of
natural languages, and that is by now sufficiently robust to offer a serious
alternative to the empiricist tradition. Like Pinker and Plotkin, I think the
new rationalism is the best story about the mind that science has found to
tell so far. But I think their version of that story is tendentious; indeed
importantly flawed. And I think the cheerful tone that they tell it in is
quite unwarranted by the amount of progress that has actually been made.
Our best scientific story about the mind is better than empiricism; but, in
204 Chapter 17
all sorts of ways, it's still not very good. Pinker quotes Chomsky's remark
that "ignorance can be divided into problems and mysteries. . . . I wrote
this book [Pinker continues] because dozens of mysteries of the mind,
from mental images to romantic love, have recently been upgraded to
problems (though there are still some mysteries too!)" (ix) Well, cheerful-
ness sells books, but what Ecclesiastes said still holds: ''The heart of the
wise is in the house of mourning."
Pinker elaborates his version of rationalism around four main ideas:
The mind is a computational system.
The mind is massively modular.
A lot of mental structure, including a lot of cognitive structure, is
innate.
A lot of mental structure, including a lot of cognitive structure, is
an evolutionary adaptation. In particular, the function of a creature's
nervous system is to abet the propagation of its genome (its selfish
genes, as one says).
Plotkin agrees with all four of these theses, though he puts less empha-
sis than Pinker does on the minds-are-computers part of the story. Both
authors take for granted that psychology should be a part of biology, and
they are both emphatic about the need for more Darwinian thinking in
cognitive science. (Plotkin quotes with approval Theodore Dobzhansky' s
dictum that "nothing in biology makes sense except in the light of evolu-
tion," amending it, however, to read "makes complete sense.") It's their
Darwinism, specifically their allegiance to a "selfish gene" account of the
phylogeny of mind, that most strikingly distinguishes Pinker and Plotkin
from a number of their rationalist colleagues (from Chomsky in particu-
lar). All this needs some looking into. I'll offer a sketch of how the four
pieces of the Pinker-Plotkin version of rationalism are connected; and, by
implication, of what an alternative rationalism might look like. I'm partic-
ularly interested in how much of the Pinker-Plotkin consensus turns on
the stuff about selfish genes, of which I don't, in fact, believe a word.
Computation
Beyond any doubt, the most important thing that has happened in cogni-
tive science was Turing's invention of the notion of mechanical ratio-
nality. Here's a quick, very informal, introduction. (Pinker provides one
that's more extensive.)
It's a remarkable fact that you can tell, just by looking at it, that any
sentence of the syntactic form P and Q ("John swims and Mary drinks," as
it might be) is true if and only if P and Q are both true. ''You can tell just
by looking" means: to see that the entailments hold, you don't have to
know anything about what either P or Q means, and you don't have
Review of Pinker and Plotkin 205
to know anything about the nonlinguistic world. This really is remarkable
since, after all, it's what they mean, together with how the nonlinguistic
world is, that decide whether either P or Q themselves are true. This
line of thought is often summarized by saying that some inferences
are rational in virtue of the syntax of the sentences that enter into them;
metaphorically, in virtue of the "shapes" of these sentences.
Turing noticed that, wherever an inference is formal in this sense, a
machine can be made to execute the inference. This is because although
machines are awful at figuring out what things mean and are not much
better at figuring out what's going on in the world, you can make them so
that they are quite good at detecting and responding to syntactic rela-
tions between sentences. Given an argument that depends just on the
syntax of the sentences that it is couched in, such a device will accept the
argument if and only if it is valid. To that extent, you can build a rational
machine. Thus, in chrysalis, the computer and all its works.
Thus too the idea that some, at least, of what makes minds rational is
their ability to perform computations on thoughts; where thoughts, like
sentences, are assumed to be syntactically structured, and where "compu-
tations" means formal operations in the manner of Turing. It's this theory
that Pinker has in mind when he claims that "thinking is a kind of compu-
tation" (21). It has proved to be a simply terrific idea. Like Truth, Beauty,
and Virtue, Rationality is a normative notion; the computational theory of
mind is the first time in all of intellectual history that a science has been
made out of one of those. If God were to stop the show now and ask us
what we've discovered about how we think, Turing's theory of computa-
tion is by far the best thing that we could offer. But.
Turing's account of computation is, in a couple of senses, local. It doesn't
look past the form of sentences to their meanings; and it assumes that the
role of thoughts in a mental process is determined entirely by their inter-
nal (syntactic) structure. And there's reason to believe that at least some
rational processes are not local in either of these respects. It may be that
wherever either semantic or global features of mental processes begin
to make their presence felt, you reach the limits of what Turing's kind
of computational rationality is able to explain. As things stand, what's
beyond these limits is not a problem but a mystery.
For example, I think it's likely that a lot of rational belief formation
turns on what philosophers call "inferences to the best explanation."
You're given what perception presents to you as currently the fact, and
you're given what memory presents to you as the beliefs that you've
formed till now. Your cognitive problem is to find and adopt whatever
new beliefs are best confirmed on balance. ''Best confirmed beliefs on
balance" means something like: the strongest and simplest relevant beliefs
that are consistent with as many of one's prior epistemic commitments as
206 Chapter 17
possible. But, as far as anybody knows, relevance, strength, simplicity,
centrality, and the like are properties, not of single sentences, but of whole
belief systems; and there's no reason at all to suppose that such global
properties of belief systems are syntactic.
In my view, the cognitive science that we've got so far has hardly
begun to face this issue. Most practitioners (Pinker and Plotkin included,
as far as I can tell) hope that it will resolve itself into lots of small, local
problems which will in tum succumb to Turing's kind of treatment.
Well, maybe; it's certainly worth the effort of continuing to try. But I'm
impressed by this consideration: Our best cognitive science is the psy-
chology of language and the psychology of perception, and (see just
below) it may well be that linguistic and perceptual processes are largely
modular, hence computationally local. Howerer, plausibly, the globality
of cognition shows up clearest in common sense reasoning. Uncoinci-
dentally, as things now stand, we don't have a theory of the psychology
of common sense reasoning that would survive serious scrutiny by an
intelligent five-year-old. Likewise, common sense is what the computers
that we know how to build egregiously don't have. I think it's likely that
we are running into the limits of what can be explained with Turing's kind
of computation; and I think we don't have any idea what to do about it.
Suffice it, anyhow, that if your notion of computation is exclusively
local, like Turing's, then your notion of mental architecture had best
be massively modular. That brings us to the second tenet of the Pinker-
Piotkin version of rationalism.
Massive modularity A module is a special purpose, more or less autono-
mous computational system. It's built to solve a very restricted class of
problems, and the information it can use to solve them with is propri-
etary. Most of the new rationalists think that at least some human cogni-
tive mechanisms are modular, aspects of language and perception being
among the classical best candidates. For example, the computations that
convert a two-dimensional array of retinal stimulations into a stable
image of a three-dimensional visual world are supposed to be largely
autonomous with respect to the rest of one's cognition. That's why many
visual illusions don't go away even if you know that they are Ulusory.
Massimo Piatelli, reviewing Plotkin's book in Nature, remarks that the
modularity of cognitive processes is "arguably ... the single most impor-
tant discovery of cognitive science." At a minimum, it's what most dis-
tinguishes our current cognitive science from its immediate precursor, the
''New Look" psychology of the 1950s that emphasized the continuity of
perception with cognition, and hence the impact of what one believes on
what one sees.
Review of Pinker and Plotkin 207
Both Pinker and Plotkin think the mind is mostly made of modules;
that's the massive modularity thesis. I want to stress how well it fits the
idea that mental computation is local. By definition, modular problem
solving works with less than everything that a creature knows. It thereby
minimizes the global cognitive effects that are the bane of Turing's kind
of computation. If the mind is massively modular, then maybe the notion
of computation that Turing gave us is, after all, the only one that cogni-
tive science needs. It would be nice to be able to believe that; Pinker and
Plotkin certainly try very hard to.
But it seems to me one can't. Eventually the mind has to integrate the
results of all those modular computations, and I don't see how there could
be a module for doing that. The moon looks bigger when it's on the hori-
zon; but I know perfectly well it's not. My visual perception module gets
fooled, but I don't. The question is: who is this 17 And by what-presum-
ably global-computational process does it use what I know about the
astronomical facts to correct the misleading appearances that my visual
perception module insists on computing? If, in short, there is a community
of computers living in my head, there had also better be somebody who
is in charge; and, by God, it had better be me. The old rationalists,
like Kant, thought that the integration of information is a lot of what's
required to tum it into knowledge. If that's right, then a cognitive science
that hasn't faced the integration problem has scarcely gotten off the
ground. Probably, modular computation doesn't explain how minds are
rational; it's just a sort of precursor. It's what you have to work through
to get a view of how horribly hard our is to understand.
Innateness Rationalists are nativists by definition; and nativism is where
cognitive science touches the real world. As both Pinker and Plotkin
rightly emphasize, the standard view in current social science-and in
what's called '1iterary theory" -takes a form of empiricism for granted:
human nature is arbitrarily plastic, and minds are social constructs. By
contrast, the evidence &om cognitive science is that a lot of what's in the
modules seems to be there innately. Pinker and Plotkin both review a fair
sample of this evidence, including some of the lovely experimental work
on infant cognition that psychologists have done in the last couple of
decades. There is also, as the linguists have been claiming for years, a lot
of indirect evidence that points to much the same conclusion: All human
languages appear to be structurally similar in profound and surprising
ways. There may be an alternative to the nativist explanation that lin-
guistic structure is genetically specified; but, if there is, nobody thus far
has had a glimpse of it.
Cultural relativism is widely held to be politically correct. If so, then
sooner or later political correctness and cognitive science are going to
208 Chapter 17
collide. Many tears will be shed and many hands will be wrung in public.
So be it. If there is a human nature, and if it is to some interesting extent
genetically determined, it is folly for humanists to ignore it. We're animals
whatever else we are; and what makes an animal well, and happy, and
sane depends a lot on what kind of animal it is. Pinker and Plotkin are
both very good on this; I commend them to you. But, for present pur-
poses, I want to examine a different aspect of their rationalism.
Psychological Darwinism Pinker and Plotkin both believe that if nativism
is the right story about cognition it follows that much of our psychology
must be, in the Darwinian sense, an evolutionary adaptation; that is,
it must be intelligible in light of evolutionary selection pressures that
shaped it. It's the nativism that makes cognitive science politically inter-
esting. But it's the inference from nativism to Darwinism that is currently
divisive within the rationalist community. Pinker and Plotkin are selling
an evolutionary approach to psychology that a lot of cognitive scientists
(myself included) aren't buying. I want to spend some time on this.
There are two standard arguments, both of which Pinker and Plotkin
endorse, that are supposed to underwrite the inference from nativism to
psychological Darwinism. The first is empirical, the second is methodo-
logical. I suspect that both are wrong headed.
The empirical argument is that, as a matter of fact, there is no way
except evolutionary selection for nature to build a complex, adaptive sys-
tem. Plotkin says that "neo-Darwinian theory [is] the central theorem of
all biology, including behavioral biology," (54) and that '1f behavior is
adaptive, then it must be the product of evolution" (53). Likewise Pinker:
''Natural selection is the only explanation we have of how complex life
can evolve ... [so] natural selection is indispensable to understanding the
human mind" (ISS).
One reply to this argument is that there is, after all, an alternative to
natural selection as the source of adaptive complexity; you could get
some by a miracle. But I'm not a creationist, nor are any of my rationalist
friends, as far as I know. Nor do we have to be, since there's another way
out of the complexity argument. This is a long story, but here's the gist:
It's common ground that the evolution of our behavior was mediated by
the evolution of our brains. So, what matters to the question of whether
the mind is an adaptation is not how complex our behavior is, but how
much change you would have to make in an ape's brain to produce the
cognitive structure of a human mind. And about this, exactly nothing is
known. That's because nothing is known about how the structure of our
minds depends on the structure of our brains. Nobody even knows which
brain structures our cognitive capacities depend on.
Review of Pinker and Plotkin 209
Unlike our minds, our brains are, by any gross measure, very similar to
those of apes. So it looks as though relatively small alterations of brain
structure must have produced very large behavioral discontinuities in the
transition &om the ancestoral apes to us. If that's right, then you don't
have to assume that cognitive complexity is shaped by the gradual action
of Darwinian selection on prehuman behavioral phenotypes. Analogies
to the evolution of organic structures, though they pervade the literature
of psychological Darwinism, actually don't cut much ice here. Make the
giraffe's neck just a little longer and you correspondingly increase, by just
a little, the animal's capacity to reach the &uit at the top of the tree. So it's
plausible, to that extent, that selection stretched giraffes' necks bit by bit.
But make an ape's brain just a little bigger (or denser, or more folded, or,
who knows, greyer) and it's anybody' s guess what happens to the crea-
ture's behavioral repertoire. Maybe the ape turns into us.
Adaptationists about the phylogeny of cognition say that it's a choice
between Darwin and God, and they like to parade as scienti6cally tough-
minded about which one of these you should pick. But that misstates the
alternatives, so don't let yourself be bullied. In fact, we don't know what
the scienti6cally reasonable view of the phylogeny of behavior is; nor will
we until we begin to understand how behavior is subserved by the brain.
And bother tough-minded anyhow; what matters is what's true.
The methodological argument &om nativism to Darwinism fares no
better. Pinker and Plotkin both hold that the (anyhow, a) proper method
of cognitive psychology is "reverse engineering." Reverse engineering is
inferring how a device must work &om, inter alia, a prior appreciation of
its function. If you don't know what a can opener is for, you are going to
have trouble Aguring out what its parts do. In the case of more complex
machines, like for example people, your chance of getting the structure
right is effectively nil if you don't know the function. Psychological
Darwinism, so the argument goes, gives us the very notion of function
that the cognitive scientist's reverse engineering of the mind requires:
To a Arst approximation, and with, to be sure, occasional exceptions, the
function of a cognitive mechanism is whatever evolution selected it for.
Without this evolutionary slant on function, cognitive science is therefore
simply in the dark. So the argument goes.
This too is a long story. But if evolution really does underwrite a
notion of function, it is a historical notion; and it's far &om clear that a
historical notion of function is what reverse engineering actually needs.
You might think, after all, that what matters in understanding the mind is
what ours do now, not what our ancestors' did some millions of years
ago. And, anyhow, the reverse engineering argument is over its head in
anachronism. As a matter of fact, lots of physiology got worked out long
210 Chapter 17
before there was a theory of evolution. That's because you don't have to
know how hands (or hearts, or eyes, or livers) evolved to make a pretty
shrewd guess about what they are for. Maybe you likewise don't have to
know how the mind evolved to make a pretty shrewd guess at what it's
for; for example, that it's to think with. No doubt, arriving at a "complete"
explanaHon of the mind by reverse engineering might require an appreci-
ation of its evolutionary history. But I don't think we should be worrying
much about complete explanations at this stage. I, for one, would settle
for the merest glimpse of what is going on.
One last point about the status of the inference &om nativism to
psychological Darwinism. If the mind is mostly a collection of innate
modules, then pretty clearly it must have evolved gradually, under selec-
tion pressure. That's because, as I remarked above, modules contain lots
of specialized information about the problem domains that they compute
in. And it really would be a miracle if all those details got into brains via a
relaHvely small, fortuitous alteraHon to some ape's neurology. To put it
the other way around, if adaptationism isn't true in psychology, it must
be that what makes our minds so clever is something pretty general;
something about their global structure. The moral is that if you aren't into
psychological Darwinism, you shouldn't be into massive modularity
either. Everything connects.
For the sake of the argument, however, let's suppose that the mind is an
adaptation after all and see where that assumpHon leads. It's a point of
definition that adaptations have to be for something. Pinker and Plotkin
both accept the "selfish gene" story about what biological adaptaHons are
for. Organic structure is (mostly) in aid of the propagation of the genes.
And so is brain structure inter alia. And so is cognitive structure, since
how the mind works depends on how the brain does. So there's a route
&om Darwinism to sociobiology; and Pinker, at least, is keen to take it.
(Plotkin seems a bit less so. He's prepared to settle for arguing that some
of the apparently egregious problems for the selfish gene theory-the
phylogeny of altruism, for example-may be less decisive than were at
first supposed. I think that settling for that is very wise of him.)
A lot of the fun of Pinker's book is his attempt to deduce human psy-
chology &om the assumption that our minds are adaptaHons for trans-
mitting our genes. His last several chapters are devoted to this, and they
range very broadly, including, so help me, one on the meaning of life.
Pinker would like to convince us that the predicHons that the selfish gene
theory makes about how our minds must be organized are independently
plausible. But this project doesn't fare well either. Prima facie, the picture
of the mind, indeed of human nature in general, that psychological
Darwinism suggests is preposterous; a sort of jumped up, down market
version of original sin.
Review of Pinker and Plotkin 211
Psychological Darwinism is a kind of conspiracy theory; that is, it
explains behavior by imputing an interest (viz., one in the proliferation of
the genome) that the agent of the behavior does not acknowledge. When
literal conspiracies are alleged, duplicity is generally part of the charge:
"He wasn't making confetti; he was shredding the evidence. He did X in
aid of Y, and then he lied about his motive." But in the kind of conspiracy
theories that psychologists like best, the motive is supposed to be in-
accessible even to the agent, who is thus perfectly sincere in denying the
imputation. In the extreme case, it hardly even is the agent to whom the
motive is attributed. Freudian explanations provide a familiar example:
What seemed to be merely Jones's slip of the tongue was the unconscious
expression of a libidinous impulse. But not Jones's libidinous impulse,
really; one that his Id had on his behalf. Likewise, for the psychological
Darwinist: What seemed to be your (after all unsurprising) interest in
your child's well-being turns out to be your genes' conspiracy to propa-
gate themselves. Not your conspiracy, but theirs.
How do you make out the case that Jones did X in aid of an interest in
Y, when Y is an interest that Jones doesn't own up to? The idea is per-
fectly familiar: You argue that X would have been the rational (reason-
able, intelligible) thing for Jones to do if Y had been his motive. Such
arguments can be very persuasive. The files Jones shredded were precisely
the ones that incriminated him; and he shredded them in the middle of the
night. What better explanation than that Jones conspired to destroy the
evidence? Likewise when the conspiracy is unconscious. Suppose that an
interest in the propagation of the genome would rationalize monogamous
families in animals whose offspring mature slowly. Well, our offspring do
mature slowly; and our species does, by and large, favor monogamous
families. So that's evidence that we favor monogamous families because
we have an interest in the propagation of our genes. Well, isn't it?
Maybe yes, maybe no; this kind of inference needs to be handled with
great care. For, often enough, where an interest in X would rationalize Y,
so too would an interest in P, Q, orR. It's reasonable of Jones to carry an
umbrella if it's raining and he wants to keep dry. But, likewise, it's reason-
able of Jones to carry an umbrella if he has in mind to return it to its
owner. Since either motivation would rationalize the way that Jones
behaved, his having behaved that way is compatible with either inter-
pretation. This is, in fact, overwhelmingly the general case: There are,
almost always, all sorts of interests that would rationalize the kinds of be-
havior that a creature is observed to produce. What's needed to make it
decisive that the creature is interested in Y is that it should produce a kind
of behavior that would be reasonable only given an interest in Y. But
such cases are vanishingly rare, since if an interest in Y would rationalize
doing X, so too would an interest in doing X. A concern to propagate
212 Chapter 17
one's genes would rationalize one's acting to promote one's children's
welfare; but so too would an interest in one's children's welfare. Not all of
one's motives could be instrumental, after all; there must be some things
that one cares for just for their own sakes. Why, indeed, mightn't there be
quite a few such things? Why mightn't one's children be among them?
The literature of psychological Darwinism is full of what appear to be
fallacies of rationalization: arguments where the evidence offered that an
interest in Y is the motive for a creature's behavior is primarily that an
interest in Y would rationalize the behavior if it were the creature's
motive. Pinker's book provides so many examples that one hardly knows
where to start. Here's Pinker on friendship: "Once you have made your-
self valuable to someone, the person becomes valuable to you. You value
him or her because if you were ever in trouble, they would have a
stake--albeit a selAsh stake--in getting you out. But now that you value
the person, they should value you even more ... because of your stake in
rescuing him or her from hard times .... This runaway process is what we
call friendship" (508-509).
And here's Pinker on why we like Action: ''Fictional narratives supply
us with a mental catalogue of the fatal conundrums we might face some-
day and the outcomes of strategies we could deploy in them. What are
the options if I were to suspect that my uncle killed my father, took his
position, and married my mother?" (543) Good question. Or what if it
turns out that, having just used the ring that I got by kidnapping a dwarf
to pay off the giants who built me my new castle, I should discover that it
is the very ring that I need in order to continue to be immortal and rule
the world? It's important to think out the options betimes, because a thing
like that could happen to anyone and you can never have too much
insurance.
At one point Pinker quotes H. L. Mencken' s wisecrack that "the most
common of all follies is to believe passionately in the palpably not true."
Quite so. I suppose it could turn out that one's interest in having friends,
or in reading Action, or in Wagner's operas, is really at heart prudential.
But the claim affronts a robust, and I should think salubrious, intuition
that there are lots and lots of things that we care about simply for them-
selves. Reductionism about this plurality of goals, when not Philistine or
cheaply cynical, often sounds simply funny. Thus the joke about the
lawyer who is offered sex by a beautiful woman. 'Well, I guess so," he
replies, ''but what's in it for mer' Does wanting a beautiful woman-or,
for that matter, just a good read-really require a further motive to
explain it? Pinker duly supplies the explanation that you wouldn't have
thought you needed: "Both sexes want a spouse who has developed nor-
mally and is free of infection .... We haven't evolved stethoscopes or
tongue depressors, but an eye for beauty does some of the same
Review of Pinker and Plotkin 213
things .... Luxuriant hair is always pleasing, possibly because ... long hair
implies a long history of good health" (483-484).
Much to his credit, Pinker does seem a bit embarrassed about some of
these consequences of his adaptationism, and he does try to duck them.
''Many people think that the theory of the selfish gene says that 'animals
try to spread their genes.' This misstates ... the theory. Animals, includ-
ing most people, know nothing about genetics and care even less. People
love their children not because they want to spread their genes (con-
sciously or unconsciously) but because they can't help it .... What is self-
ish is not the real motives of the person but the metaphorical motives of
the genes that built the person. Genes 'try' to spread themselves [sic] by
wiring animals' brains so the animals love their kin ... and then the[y] get
out of the way" (401). This version sounds a lot more plausible; strictly
speaking, nobody has as a motive ("conscious or unconscious") the pro-
liferation of genes after all. Not animals, and not genes either. The only
motives there really are, are the ones that everybody knows about; of
which love of novels, or women, or kin are presumably a few among a
multitude.
Right on. But, pace Pinker, this reasonable view is not available to a
psychological Darwinist. For to say that the genes 'wire animals' brains
so the animals love their kin" and to stop there is to say only that loving
their kin is innate in these animals. That reduces psychological Darwinism
to mere nativism; which, as I remarked above, is common ground to all of
us rationalists. The difference between Darwinism and mere nativism is
the claim that a creature's innate psychological' traits are adaptations; that
is, that their role in the propagation of the genes is what they're for and is
what makes their structure intelligible. Take the adaptationism away from
a psychological Darwinist, and he has nobody left to argue with except
empiricists.
It is, then, adaptationism that makes Pinker's and Plotkin's kind of
rationalism special. Does this argument between rationalists really matter?
Nativism itself clearly does; everybody cares about human nature. But I
have fussed a lot about the difference between nativism and Darwinism,
and you might reasonably want to know why anyone should care about
that.
For one thing, nativism says there is a human nature, but it's the adap-
tationism that implies the account of human nature that sociobiologists
endorse. If, like me, you find that account grotesquely implausible, it's
perhaps the adaptationism rather than the nativism that you ought to
consider throwing overboard. Pinker remarks that "people who study the
mind would rather not have to think about how it evolved because it
would make a hash of cherished theories .... When advised that [their]
claims are evolutionarily implausible, they attack the theory of evolution
214 Chapter 17
rather than rethinking the claim" (165). I think this is exadly right, though
the formulation is a bit tendentious. We know-anyhow we think we
know-a lot about ourselves that doesn't seem to square with the theory
that our minds are adaptations for spreading our genes. The question may
well come down to which theory we should give up. Well, as far as I can
tell, if you take away the bad argument that turns on complexity, and
the bad argument &om reverse engineering, and the bad arguments that
depend on committing the rationalization fallacy, and the atrociously bad
arguments that depend on preempting what's to count as the "scientific"
(and/or the biological) world view, the dired evidence for psychological
Darwinism is very slim indeed. In particular, it's arguably much worse
than the circumstantial evidence for our intuitive, pluralistic theory of
human nature. It is, after all, our intuitive pluralism that we use to get
along with one another. And I have the impression that, by and large, it
works pretty well.
So things could be worse, sociobiologists to the contrary not with-
standing; I do hope you find that consoling. Have a nice millenium.
References
Allen. C. and Bekoff, M. 1995. "Biological Function, Adaptation, and Natural Design,"
Philosophy of Science 62 (4): 609-622.
Berlin, B. and Kay, P. 1969. &sic Color Terms: Their Universality mul Evolution. Berkeley:
University of California Press.
Block. N. 1978. ''Troubles with Functionalism," in C. W. Savage (ed.), Minnesot11 Studies in the
Philosophy of Science, Vol IX. Minneapolis, MN: University of Minnesota Press.
Block, N. 1993. ''Holism, Hyper-Analyticity, and Hyper-Compositionality," Mirul mul Llm-
guage 8: 1-26.
Block, N. and Fodor, J. 1972. ''What Psychological States Are Not," Philosophiad Review 81:
159-181.
Carruthers, P. 1996. lAnguRge, Thought, Rrul Consciousness. Cambridge: Cambridge University
Press.
Churchland, Patricia. 1987. ''Epistemology in the Age of Neuoscience," ]ounud of Philosophy
84 (10): 544-555.
Churchland, Paul. 1995. The Engine of Rttuon, The Stilt of the Soul. Cambridge, MA: MIT
Press.
Davidson, D. 1980. ''Mental Events," in D. Davidson, f.ssllys on Actions Rnll Events. Oxford:
Clarendon Press.
Dawkins, R. 1976. The Selfish Gene. Oxford: Oxford University Press.
Dawkins, R. 1996. Climbing Mount Improbable. New York: Viking.
Dennett, D. C. 1978a. ''Intentional Systems," in D. C. Dennett, Bminstomrs, q.v.
Dennett, D. C. 1978b. Brainstorms. Cambridge, MA: Bradford Books/MIT Press.
Dennett, D. C. 1990. ''Teaching an Old Dog New Tricks," BehRviOf'tll Rrul Br11in Sciences
13:76-77.
Dennett, D. C. 1991. ''Real Patterns," ]ounuJ of Philosophy 87: 27-51.
Dennett, D. C. 1995. Dllrwin's Dtmgerous Idell. New York: Simon and Schuster.
Devitt, M. 1996. Coming to Our Senses. Cambridge: Cambridge University Press.
Eldredge, N. 1995. Reinventing Dmwin. New York: John Wiley.
Elman, J. et al. 1996. Rethinking Innateness: A Connectionist Perspective on Dtwlopmmt.
Cambridge, MA: MIT Press.
Fodor, J. 1974. "Special Sciences," Synthese 28: 97-115.
Fodor, J. 1990. "A Theory of Content, Part I," in J. Fodor, A Theory of Content Rrul Other
fssllys. Cambridge, MA: MIT Press.
Fodor, J. 1991. ''You Can Fool All of the People Some of the Time, Everything Else Being
Equal: Hedged Laws in Psychological Explanations," Mirul100 (1): 19-34.
Fodor, J. and Lepore, E. 1992. Holism: A Shopper's GuUle. Oxford: Blackwell.
Fodor, J. and Pylyshyn, Z. 1988. "Connectionism and Cognitive Architecture: A Critical
Analysis," Cognition 28: 3-71.
216 References
Gallistel, C. 1970. The OrganiZJition of Learning. Cambridge, MA: MIT Press.,
Gleitman, H. 1995. Psychology, fourth edition. New York: W. W. Norton & Co.
Gould, S. J. and Lewontin, R. C. 1979. "The Spandrels of San Marco and the Panglossian
Paradigm: A Critique of the Adaptationist Program," Proceedings of the Royal Society of
London B 205: 581-598.
Hempel, C. G. 1965. ''The Logic of Functional Analysis," in C. G. Hempel, Aspects of Scientific
Explan.ation. New York: Macmillan.
Kamp, H. and Partee, B. 1995. ''Prototype Theory and Compositionality," Cognition 57: 129-
181.
Karmiloff-Smith, A. 1992. Beyond Modularity: A Developmental Perspective on Cognitive Science.
Cambridge, MA: MIT Press.
Kim, J. 1992. "Multiple Realization and the Metaphysics of Reduction," Philosophy and
Phenomenological Research 52: 1-26.
Kim, J. 1993. Supervenience and Mind. Cambridge: Cambridge University Press.
Loewer, B. and Rey, G. (eds). 1991. Meaning in Mind: Fodor and His Critics. Oxford: Black-
well.
Macdonald, C. and Macdonald, G. (eds.). 1995. Connectionism. Cambridge, MA: Blackwell.
Mar-cus, G. F., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J., and Xu, F. 1992. "Over-
regularization in Language Acquisition," Monographs of the Society for Research in
Child Development 57: 81-85.
McCauley, R. (ed.). 1995. The Churchlands and Their Critics. Cambridge, MA: Blackwell.
McDowell, J. 1994. Mind and World. Cambridge, MA: Harvard University Press.
Millikan, R. 1984. Language, Thought, and Other Biological Categories. Cambridge, MA: MIT
Press.
Mithen, S. 1996. The Prehistory of Mind. Thames and Hudson.
Moore, A. W. 1997. Points of View. Oxford: Oxford University Press.
Orr, H. A. 1996. "Dennett's Dangerous Idea," Evolution SO (1): 467-472.
Peacocke, C. 1992. A Study of Concepts. Cambridge, MA: MIT Press.
Pinker, S. 1997. How the Mind Works. New York: Norton.
Plantinga, A. (forthcoming). "Naturalism Defeated." Unpublished manuscript, University of
Notre Dame.
Plotkin, H. 1997. Evolution in Mind. London: Alan Lane.
Putnam, H. 1967. ''The Nature of Mental States," in D. Rosenthal (ed.), q.v.
Pylyshyn, Z. 1984. Computation and Cognition: Toward a Foundation for Cognitive Science.
Cambridge, MA: MIT Press.
Rosch, E. H. 1973. ''Natural Categories," Cognitive Psychology 4: 328-350.
Rosenthal, D. (ed.). 1991. The Nature of Mind. Oxford: Oxford University Press.
Savin, H. and Bever, T. G. 1970. ''The Nonperceptual Reality of the Phoneme," journal of
Verbal Learning and Verbal Behavior 9: 295-302.
Schiffer, S. 1991. "Ceteris Paribus Laws," Mind 100 (1): 1-17.
Smolensky, P. 1987. "The Constituent Structure of Mental States: A Reply to Fodor and
Pylyshyn," Southern journal of Philosophy 26: 137-160.
Smolensky, P. 1988a. "On the Proper Treatment of Connectionism," Behavioral and Brain
Sciences 11: 1-23.
Smolensky, P. 1990. "Tensor Product Variable Binding and the Representation of Structure
in Connectionist Systems," Arlifidallntelligence 46, 159-216.
Smolenksy, P. 199Sa. "Connectionism, Constituency, and the Language of Thought," in
Macdonald and Macdonald, q.v.
Smolensky, P. 199Sb. ''Reply: Constituent Structure and Explanation in an Integrated
Connectionist/Symbolic Cognitive Architecture," in Macdonald and Macdonald, q.v.
Sober, E. 1984. The Nature of Selection. Cambridge, MA: MIT Press.
Index
Accessibility (of modularized information),
121. See 11lso Demodularization (of mental
processes)
Adaptationism. See Darwin, C. (Darwinism)
Agglomerative principle, 45-46, 61n
Allen, C., 183
Altruism, 168
Analytic/synthetic distinction, 32-33, SO,
11-12
Aristotle, be, 88
Associationism. See Empiricism
Atomism (of meaning), 63, 69-72
Bekoff, M., 183
Berkeley, G., 193
Berlin, B., 47n
Bever, T. G., 133
Block. N., 10, 23, 46, 6ln, 69
Brecht, B., x
Brentano, F., 154, 183, 184
Bridge laws, 16-17
Carruthers, P., 63-73
Chomsky, N., 27, 65, 128, 141, 153, 154,
ISS, 164-165,175,203,204
Churchland, Patricia. 184, 189-192, 195,
2oo-20l
Churchland, Paul, 83-89, 184
Cognitive neuroscience, 83-89
Communication, 67
Compositionality, 36-46, So-60
Oassical treatment of, 92-25
strong and wealc, 91-111
Concept possession, 28, 36-42, SO, 11
Concepts, 27-34
constitution of, 31-33, SO
good instances of, 38-39,42-43, 51, ss-
57
Conceptual analysis, 29-30, 33, 37-38
Connectionism, 83-89, 91-111, 139-141,
143-151
Consciousness
and "higher order'' thought, 73
and thought, 63
Constituent structure (of mental representa-
tions), 91-111
cummings, e. e., 9
Darwin, C. (Darwinism), 23, 158, 163-169,
1 7 1 - 1 8 ~ 186,189-202,208-214
Davidson, D., 3, 4, 10, 23, 27, 11
Dawkins, R., 163-169, 173
Demodularization (of mental processes),
13o-134
Dennett, D., 23, 27, 111-187, 200
Derrida, J,, 11
Descartes, R., ix, 7, 8, 29, 44, 146, 202
Devitt, M., 70
Dewey, J,, 3
Disjunctive kinds, 11-13
Distributed representations, 91-111
Dobzhansky, T., 204
Domain specificity (of modular processes),
1 2 7 - 1 4 ~ 146, 154-155
Einstein, A, 88
Eldredge, N., 24n, 166, 174-175
Elman, J,, 128, 129, 142n, 143-151
Empiricism, 27, 35, 40, 51, 53, 57, 58, 148-
151, 153, 203
Encapsulation (of mental processes), 127,
154-160. See 11lso General intelligence
Epigenesis, 128, 130, 141n, 146-148
Epistemology, 4-8, 43
Evolution, theory of. See Darwin, C.
(Darwinism)
218 Index
Fodor,)., 10, 12, 16, 23, 46n, 47n, 63, 91, 95,
128, 178, 180, 186
Forster, K., 64
Frege, G., 27, 29, 65
Freud, S., 153, 211
Functionalism, 9-22
in semantics, 69-72
Gall F., 153
General intelligence, 157-160
Gleitman. H., 89
Globality (of mental processes), 205-
206
Goodman, N., 77
Gould, S., 172, 174, 175, 199
Grice, H. P., 63, 66-68
Hegel, G. W. F., 3
Heidegger, M., 27, 43
Hempel, C., 178-179,202
HUI climbing, 163-165
Horwich, P., 46, 61n
Hume, D., ix, 29, 64, 85, 148, 149, 203
Idealism, 6
Indexicals, 76
Innateness of mental states. Su Nativism
Intentional concepts, 42
Intentionality (and teleology), 176-186
Introspection, 64-65
Justification, 5-8
Kamp, H., 47n
Kant, 1., ix. 3, 64, 75, 77, 78, 150, 153,
207
Kanniloff-Smith, A, 127-142, 157
Kay, P., 47n
Kim, )., 9-22
Kripke, S., 5
Kuhn, T., 77, 78, 79
Laws (of psychology), 1-8, 9-22
Leamability (of the lexicon), 53-60
Leibniz. G., 150
Lepore, E., 23, 46n, 47n, 55
Levine,)., 23
Lewontin, R., 17 4, 199
Loar, B., 61n
Locke, )., 203
Loewer, B., 23, 180
Marcus, G., 142n
Mates, B., 32
McDowell )., 3-8
McLaughlin, B., x
Meaning (metaphysics of), 171-184
Menken. H. L., 212
Mentalese,63-64,66-73
and natural languages, 67
MiD, ). S., 29
MiUikan, R., 179-180, 186, 187n
Mithen. S., 153-159
Modularity (of mental processes), 127-142,
153-160, 203-214
"massive modularity" thesis, 206-207
Modularization (of mental processes), 128,
129
Moore, A W., 75-78
Mother Nature, 178 180-185, 186, 187n
Multiple realization (of functional states), 9-
22
Nativism, 128-142, 143-151, 207-208
Naturalism, 3-4, 6
Neural networks, 83-89
Neural plasticity, 129-130,.144-145
Newton, I., ix
Nietzsche, F., x
Objectivity, 75-78
Orr, H. A, 174
Partee, B., 47n
Peacocke, C., x. 27-34, 46, 61n
Peirce, C., 43
Perception, 5-6
Phenotype, 166
Piaget, )., 27, 141
PiateUi, M., 206
Pinker, S., 203-214
Plantinga, A, 189-202
Plato, v, ix
Plotkin, H., 203-214
Pluralism, 4
PoweU, A, 156
Pragmatism, 43-44
Private language argument, 68
Procedural representation, 135-13 7
Productivity, 37-38, 41, 54
Projectibility, 9-11, 13
Proust, M., 75
Punctate equilibrium, 165
Putnam, H .. 3, 16, 32, 72, 77, 189
Pylyshyn. z .. 54, 91, 95
Quine, w. v. 0., 3, 27, 32, 77, 127, 184
Ramsey, F., 79
Rationalism, 148-151, 153-154, 203-214
Rationallty,4-8,33,204-206
Reductionism, 3-5, 9-22, 173, 185-186
Re8exes, 158-159
Representational redescription (of
modularized infonnation), 128, 134-
142
''Reverse engineering" (of cognitive
mechanisms), 209-210
Rey, G., 180
Rorty, R., 54
Rosch, E .. 47n
Rosenthal, D., 16
RusseU, B., 17
Ryle, G.,3
Saltation, 165, 167
Santayana, G., 151
76
Savin, H., 133
Schiffer, S., 23, 44-46, 47n, 61n

Searle, }., 186
"Selection for," 179-180, 184-185, 187n
Semantics
holism of, 69-72, 77
infonnationaL 54, 63, 185
of natural languages, 68
34n, 35-36, 41, 43, 198-199
Skinner, B. F., 164-165
Smolensky, P., 91-111
Sober, E., 187n
Spedal sdences, 9-22
Stein, H., 23
Strawson, G., 46
Systematicity, 37-38,41, 54,91-111
Teleology, 168-169, 171-187
Tenability (of gramman), 53-61
Transcendental argwnents, 27, 324n, 53-
54
Transcendental 77-78
Turing, A (Turing machines), 85, 204-206
Vector coding, 87-89, 95-111
Index 219
Wagner, R., 212
WhorE, B., 76